1 /*
2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "gc/shared/barrierSet.hpp"
26 #include "gc/shared/c2/barrierSetC2.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "memory/resourceArea.hpp"
29 #include "opto/addnode.hpp"
30 #include "opto/block.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/castnode.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/idealGraphPrinter.hpp"
35 #include "opto/loopnode.hpp"
36 #include "opto/machnode.hpp"
37 #include "opto/opcodes.hpp"
38 #include "opto/phaseX.hpp"
39 #include "opto/regalloc.hpp"
40 #include "opto/rootnode.hpp"
41 #include "utilities/macros.hpp"
42 #include "utilities/powerOfTwo.hpp"
43
44 //=============================================================================
45 #define NODE_HASH_MINIMUM_SIZE 255
46
47 //------------------------------NodeHash---------------------------------------
48 NodeHash::NodeHash(Arena *arena, uint est_max_size) :
49 _a(arena),
50 _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ),
51 _inserts(0), _insert_limit( insert_limit() ),
52 _table( NEW_ARENA_ARRAY( _a , Node* , _max ) )
53 #ifndef PRODUCT
54 , _grows(0),_look_probes(0), _lookup_hits(0), _lookup_misses(0),
55 _insert_probes(0), _delete_probes(0), _delete_hits(0), _delete_misses(0),
56 _total_inserts(0), _total_insert_probes(0)
57 #endif
58 {
59 // _sentinel must be in the current node space
60 _sentinel = new ProjNode(nullptr, TypeFunc::Control);
61 memset(_table,0,sizeof(Node*)*_max);
62 }
63
64 //------------------------------hash_find--------------------------------------
65 // Find in hash table
66 Node *NodeHash::hash_find( const Node *n ) {
67 // ((Node*)n)->set_hash( n->hash() );
68 uint hash = n->hash();
69 if (hash == Node::NO_HASH) {
70 NOT_PRODUCT( _lookup_misses++ );
71 return nullptr;
72 }
73 uint key = hash & (_max-1);
74 uint stride = key | 0x01;
75 NOT_PRODUCT( _look_probes++ );
76 Node *k = _table[key]; // Get hashed value
77 if( !k ) { // ?Miss?
78 NOT_PRODUCT( _lookup_misses++ );
79 return nullptr; // Miss!
80 }
81
82 int op = n->Opcode();
83 uint req = n->req();
84 while( 1 ) { // While probing hash table
85 if( k->req() == req && // Same count of inputs
86 k->Opcode() == op ) { // Same Opcode
87 for( uint i=0; i<req; i++ )
88 if( n->in(i)!=k->in(i)) // Different inputs?
89 goto collision; // "goto" is a speed hack...
90 if( n->cmp(*k) ) { // Check for any special bits
91 NOT_PRODUCT( _lookup_hits++ );
92 return k; // Hit!
93 }
94 }
95 collision:
96 NOT_PRODUCT( _look_probes++ );
97 key = (key + stride/*7*/) & (_max-1); // Stride through table with relative prime
98 k = _table[key]; // Get hashed value
99 if( !k ) { // ?Miss?
100 NOT_PRODUCT( _lookup_misses++ );
101 return nullptr; // Miss!
102 }
103 }
104 ShouldNotReachHere();
105 return nullptr;
106 }
107
108 //------------------------------hash_find_insert-------------------------------
109 // Find in hash table, insert if not already present
110 // Used to preserve unique entries in hash table
111 Node *NodeHash::hash_find_insert( Node *n ) {
112 // n->set_hash( );
113 uint hash = n->hash();
114 if (hash == Node::NO_HASH) {
115 NOT_PRODUCT( _lookup_misses++ );
116 return nullptr;
117 }
118 uint key = hash & (_max-1);
119 uint stride = key | 0x01; // stride must be relatively prime to table siz
120 uint first_sentinel = 0; // replace a sentinel if seen.
121 NOT_PRODUCT( _look_probes++ );
122 Node *k = _table[key]; // Get hashed value
123 if( !k ) { // ?Miss?
124 NOT_PRODUCT( _lookup_misses++ );
125 _table[key] = n; // Insert into table!
126 DEBUG_ONLY(n->enter_hash_lock()); // Lock down the node while in the table.
127 check_grow(); // Grow table if insert hit limit
128 return nullptr; // Miss!
129 }
130 else if( k == _sentinel ) {
131 first_sentinel = key; // Can insert here
132 }
133
134 int op = n->Opcode();
135 uint req = n->req();
136 while( 1 ) { // While probing hash table
137 if( k->req() == req && // Same count of inputs
138 k->Opcode() == op ) { // Same Opcode
139 for( uint i=0; i<req; i++ )
140 if( n->in(i)!=k->in(i)) // Different inputs?
141 goto collision; // "goto" is a speed hack...
142 if( n->cmp(*k) ) { // Check for any special bits
143 NOT_PRODUCT( _lookup_hits++ );
144 return k; // Hit!
145 }
146 }
147 collision:
148 NOT_PRODUCT( _look_probes++ );
149 key = (key + stride) & (_max-1); // Stride through table w/ relative prime
150 k = _table[key]; // Get hashed value
151 if( !k ) { // ?Miss?
152 NOT_PRODUCT( _lookup_misses++ );
153 key = (first_sentinel == 0) ? key : first_sentinel; // ?saw sentinel?
154 _table[key] = n; // Insert into table!
155 DEBUG_ONLY(n->enter_hash_lock()); // Lock down the node while in the table.
156 check_grow(); // Grow table if insert hit limit
157 return nullptr; // Miss!
158 }
159 else if( first_sentinel == 0 && k == _sentinel ) {
160 first_sentinel = key; // Can insert here
161 }
162
163 }
164 ShouldNotReachHere();
165 return nullptr;
166 }
167
168 //------------------------------hash_insert------------------------------------
169 // Insert into hash table
170 void NodeHash::hash_insert( Node *n ) {
171 // // "conflict" comments -- print nodes that conflict
172 // bool conflict = false;
173 // n->set_hash();
174 uint hash = n->hash();
175 if (hash == Node::NO_HASH) {
176 return;
177 }
178 check_grow();
179 uint key = hash & (_max-1);
180 uint stride = key | 0x01;
181
182 while( 1 ) { // While probing hash table
183 NOT_PRODUCT( _insert_probes++ );
184 Node *k = _table[key]; // Get hashed value
185 if( !k || (k == _sentinel) ) break; // Found a slot
186 assert( k != n, "already inserted" );
187 // if( PrintCompilation && PrintOptoStatistics && Verbose ) { tty->print(" conflict: "); k->dump(); conflict = true; }
188 key = (key + stride) & (_max-1); // Stride through table w/ relative prime
189 }
190 _table[key] = n; // Insert into table!
191 DEBUG_ONLY(n->enter_hash_lock()); // Lock down the node while in the table.
192 // if( conflict ) { n->dump(); }
193 }
194
195 //------------------------------hash_delete------------------------------------
196 // Replace in hash table with sentinel
197 bool NodeHash::hash_delete( const Node *n ) {
198 Node *k;
199 uint hash = n->hash();
200 if (hash == Node::NO_HASH) {
201 NOT_PRODUCT( _delete_misses++ );
202 return false;
203 }
204 uint key = hash & (_max-1);
205 uint stride = key | 0x01;
206 DEBUG_ONLY( uint counter = 0; );
207 for( ; /* (k != nullptr) && (k != _sentinel) */; ) {
208 DEBUG_ONLY( counter++ );
209 NOT_PRODUCT( _delete_probes++ );
210 k = _table[key]; // Get hashed value
211 if( !k ) { // Miss?
212 NOT_PRODUCT( _delete_misses++ );
213 return false; // Miss! Not in chain
214 }
215 else if( n == k ) {
216 NOT_PRODUCT( _delete_hits++ );
217 _table[key] = _sentinel; // Hit! Label as deleted entry
218 DEBUG_ONLY(((Node*)n)->exit_hash_lock()); // Unlock the node upon removal from table.
219 return true;
220 }
221 else {
222 // collision: move through table with prime offset
223 key = (key + stride/*7*/) & (_max-1);
224 assert( counter <= _insert_limit, "Cycle in hash-table");
225 }
226 }
227 ShouldNotReachHere();
228 return false;
229 }
230
231 //------------------------------round_up---------------------------------------
232 // Round up to nearest power of 2
233 uint NodeHash::round_up(uint x) {
234 x += (x >> 2); // Add 25% slop
235 return MAX2(16U, round_up_power_of_2(x));
236 }
237
238 //------------------------------grow-------------------------------------------
239 // Grow _table to next power of 2 and insert old entries
240 void NodeHash::grow() {
241 // Record old state
242 uint old_max = _max;
243 Node **old_table = _table;
244 // Construct new table with twice the space
245 #ifndef PRODUCT
246 _grows++;
247 _total_inserts += _inserts;
248 _total_insert_probes += _insert_probes;
249 _insert_probes = 0;
250 #endif
251 _inserts = 0;
252 _max = _max << 1;
253 _table = NEW_ARENA_ARRAY( _a , Node* , _max ); // (Node**)_a->Amalloc( _max * sizeof(Node*) );
254 memset(_table,0,sizeof(Node*)*_max);
255 _insert_limit = insert_limit();
256 // Insert old entries into the new table
257 for( uint i = 0; i < old_max; i++ ) {
258 Node *m = *old_table++;
259 if( !m || m == _sentinel ) continue;
260 DEBUG_ONLY(m->exit_hash_lock()); // Unlock the node upon removal from old table.
261 hash_insert(m);
262 }
263 }
264
265 //------------------------------clear------------------------------------------
266 // Clear all entries in _table to null but keep storage
267 void NodeHash::clear() {
268 #ifdef ASSERT
269 // Unlock all nodes upon removal from table.
270 for (uint i = 0; i < _max; i++) {
271 Node* n = _table[i];
272 if (!n || n == _sentinel) continue;
273 n->exit_hash_lock();
274 }
275 #endif
276
277 memset( _table, 0, _max * sizeof(Node*) );
278 }
279
280 //-----------------------remove_useless_nodes----------------------------------
281 // Remove useless nodes from value table,
282 // implementation does not depend on hash function
283 void NodeHash::remove_useless_nodes(VectorSet &useful) {
284
285 // Dead nodes in the hash table inherited from GVN should not replace
286 // existing nodes, remove dead nodes.
287 uint max = size();
288 Node *sentinel_node = sentinel();
289 for( uint i = 0; i < max; ++i ) {
290 Node *n = at(i);
291 if(n != nullptr && n != sentinel_node && !useful.test(n->_idx)) {
292 DEBUG_ONLY(n->exit_hash_lock()); // Unlock the node when removed
293 _table[i] = sentinel_node; // Replace with placeholder
294 }
295 }
296 }
297
298
299 void NodeHash::check_no_speculative_types() {
300 #ifdef ASSERT
301 uint max = size();
302 Unique_Node_List live_nodes;
303 Compile::current()->identify_useful_nodes(live_nodes);
304 Node *sentinel_node = sentinel();
305 for (uint i = 0; i < max; ++i) {
306 Node *n = at(i);
307 if (n != nullptr &&
308 n != sentinel_node &&
309 n->is_Type() &&
310 live_nodes.member(n)) {
311 TypeNode* tn = n->as_Type();
312 const Type* t = tn->type();
313 const Type* t_no_spec = t->remove_speculative();
314 assert(t == t_no_spec, "dead node in hash table or missed node during speculative cleanup");
315 }
316 }
317 #endif
318 }
319
320 #ifndef PRODUCT
321 //------------------------------dump-------------------------------------------
322 // Dump statistics for the hash table
323 void NodeHash::dump() {
324 _total_inserts += _inserts;
325 _total_insert_probes += _insert_probes;
326 if (PrintCompilation && PrintOptoStatistics && Verbose && (_inserts > 0)) {
327 if (WizardMode) {
328 for (uint i=0; i<_max; i++) {
329 if (_table[i])
330 tty->print("%d/%d/%d ",i,_table[i]->hash()&(_max-1),_table[i]->_idx);
331 }
332 }
333 tty->print("\nGVN Hash stats: %d grows to %d max_size\n", _grows, _max);
334 tty->print(" %d/%d (%8.1f%% full)\n", _inserts, _max, (double)_inserts/_max*100.0);
335 tty->print(" %dp/(%dh+%dm) (%8.2f probes/lookup)\n", _look_probes, _lookup_hits, _lookup_misses, (double)_look_probes/(_lookup_hits+_lookup_misses));
336 tty->print(" %dp/%di (%8.2f probes/insert)\n", _total_insert_probes, _total_inserts, (double)_total_insert_probes/_total_inserts);
337 // sentinels increase lookup cost, but not insert cost
338 assert((_lookup_misses+_lookup_hits)*4+100 >= _look_probes, "bad hash function");
339 assert( _inserts+(_inserts>>3) < _max, "table too full" );
340 assert( _inserts*3+100 >= _insert_probes, "bad hash function" );
341 }
342 }
343
344 Node *NodeHash::find_index(uint idx) { // For debugging
345 // Find an entry by its index value
346 for( uint i = 0; i < _max; i++ ) {
347 Node *m = _table[i];
348 if( !m || m == _sentinel ) continue;
349 if( m->_idx == (uint)idx ) return m;
350 }
351 return nullptr;
352 }
353 #endif
354
355 #ifdef ASSERT
356 NodeHash::~NodeHash() {
357 // Unlock all nodes upon destruction of table.
358 if (_table != (Node**)badAddress) clear();
359 }
360 #endif
361
362
363 //=============================================================================
364 //------------------------------PhaseRemoveUseless-----------------------------
365 // 1) Use a breadthfirst walk to collect useful nodes reachable from root.
366 PhaseRemoveUseless::PhaseRemoveUseless(PhaseGVN* gvn, Unique_Node_List& worklist, PhaseNumber phase_num) : Phase(phase_num) {
367 C->print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
368 // Implementation requires an edge from root to each SafePointNode
369 // at a backward branch. Inserted in add_safepoint().
370
371 // Identify nodes that are reachable from below, useful.
372 C->identify_useful_nodes(_useful);
373 // Update dead node list
374 C->update_dead_node_list(_useful);
375
376 // Remove all useless nodes from PhaseValues' recorded types
377 // Must be done before disconnecting nodes to preserve hash-table-invariant
378 gvn->remove_useless_nodes(_useful.member_set());
379
380 // Remove all useless nodes from future worklist
381 worklist.remove_useless_nodes(_useful.member_set());
382
383 // Disconnect 'useless' nodes that are adjacent to useful nodes
384 C->disconnect_useless_nodes(_useful, worklist);
385 }
386
387 //=============================================================================
388 //------------------------------PhaseRenumberLive------------------------------
389 // First, remove useless nodes (equivalent to identifying live nodes).
390 // Then, renumber live nodes.
391 //
392 // The set of live nodes is returned by PhaseRemoveUseless in the _useful structure.
393 // If the number of live nodes is 'x' (where 'x' == _useful.size()), then the
394 // PhaseRenumberLive updates the node ID of each node (the _idx field) with a unique
395 // value in the range [0, x).
396 //
397 // At the end of the PhaseRenumberLive phase, the compiler's count of unique nodes is
398 // updated to 'x' and the list of dead nodes is reset (as there are no dead nodes).
399 //
400 // The PhaseRenumberLive phase updates two data structures with the new node IDs.
401 // (1) The "worklist" is "C->igvn_worklist()", which is to collect which nodes need to
402 // be processed by IGVN after removal of the useless nodes.
403 // (2) Type information "gvn->types()" (same as "C->types()") maps every node ID to
404 // the node's type. The mapping is updated to use the new node IDs as well. We
405 // create a new map, and swap it with the old one.
406 //
407 // Other data structures used by the compiler are not updated. The hash table for value
408 // numbering ("C->node_hash()", referenced by PhaseValue::_table) is not updated because
409 // computing the hash values is not based on node IDs.
410 PhaseRenumberLive::PhaseRenumberLive(PhaseGVN* gvn,
411 Unique_Node_List& worklist,
412 PhaseNumber phase_num) :
413 PhaseRemoveUseless(gvn, worklist, Remove_Useless_And_Renumber_Live),
414 _new_type_array(C->comp_arena()),
415 _old2new_map(C->unique(), C->unique(), -1),
416 _is_pass_finished(false),
417 _live_node_count(C->live_nodes())
418 {
419 assert(RenumberLiveNodes, "RenumberLiveNodes must be set to true for node renumbering to take place");
420 assert(C->live_nodes() == _useful.size(), "the number of live nodes must match the number of useful nodes");
421 assert(_delayed.size() == 0, "should be empty");
422 assert(&worklist == C->igvn_worklist(), "reference still same as the one from Compile");
423 assert(&gvn->types() == C->types(), "reference still same as that from Compile");
424
425 GrowableArray<Node_Notes*>* old_node_note_array = C->node_note_array();
426 if (old_node_note_array != nullptr) {
427 int new_size = (_useful.size() >> 8) + 1; // The node note array uses blocks, see C->_log2_node_notes_block_size
428 new_size = MAX2(8, new_size);
429 C->set_node_note_array(new (C->comp_arena()) GrowableArray<Node_Notes*> (C->comp_arena(), new_size, 0, nullptr));
430 C->grow_node_notes(C->node_note_array(), new_size);
431 }
432
433 assert(worklist.is_subset_of(_useful), "only useful nodes should still be in the worklist");
434
435 // Iterate over the set of live nodes.
436 for (uint current_idx = 0; current_idx < _useful.size(); current_idx++) {
437 Node* n = _useful.at(current_idx);
438
439 const Type* type = gvn->type_or_null(n);
440 _new_type_array.map(current_idx, type);
441
442 assert(_old2new_map.at(n->_idx) == -1, "already seen");
443 _old2new_map.at_put(n->_idx, current_idx);
444
445 if (old_node_note_array != nullptr) {
446 Node_Notes* nn = C->locate_node_notes(old_node_note_array, n->_idx);
447 C->set_node_notes_at(current_idx, nn);
448 }
449
450 n->set_idx(current_idx); // Update node ID.
451
452 if (update_embedded_ids(n) < 0) {
453 _delayed.push(n); // has embedded IDs; handle later
454 }
455 }
456
457 // VectorSet in Unique_Node_Set must be recomputed, since IDs have changed.
458 worklist.recompute_idx_set();
459
460 assert(_live_node_count == _useful.size(), "all live nodes must be processed");
461
462 _is_pass_finished = true; // pass finished; safe to process delayed updates
463
464 while (_delayed.size() > 0) {
465 Node* n = _delayed.pop();
466 int no_of_updates = update_embedded_ids(n);
467 assert(no_of_updates > 0, "should be updated");
468 }
469
470 // Replace the compiler's type information with the updated type information.
471 gvn->types().swap(_new_type_array);
472
473 // Update the unique node count of the compilation to the number of currently live nodes.
474 C->set_unique(_live_node_count);
475
476 // Set the dead node count to 0 and reset dead node list.
477 C->reset_dead_node_list();
478 }
479
480 int PhaseRenumberLive::new_index(int old_idx) {
481 assert(_is_pass_finished, "not finished");
482 if (_old2new_map.at(old_idx) == -1) { // absent
483 // Allocate a placeholder to preserve uniqueness
484 _old2new_map.at_put(old_idx, _live_node_count);
485 _live_node_count++;
486 }
487 return _old2new_map.at(old_idx);
488 }
489
490 int PhaseRenumberLive::update_embedded_ids(Node* n) {
491 int no_of_updates = 0;
492 if (n->is_Phi()) {
493 PhiNode* phi = n->as_Phi();
494 if (phi->_inst_id != -1) {
495 if (!_is_pass_finished) {
496 return -1; // delay
497 }
498 int new_idx = new_index(phi->_inst_id);
499 assert(new_idx != -1, "");
500 phi->_inst_id = new_idx;
501 no_of_updates++;
502 }
503 if (phi->_inst_mem_id != -1) {
504 if (!_is_pass_finished) {
505 return -1; // delay
506 }
507 int new_idx = new_index(phi->_inst_mem_id);
508 assert(new_idx != -1, "");
509 phi->_inst_mem_id = new_idx;
510 no_of_updates++;
511 }
512 }
513
514 const Type* type = _new_type_array.fast_lookup(n->_idx);
515 if (type != nullptr && type->isa_oopptr() && type->is_oopptr()->is_known_instance()) {
516 if (!_is_pass_finished) {
517 return -1; // delay
518 }
519 int old_idx = type->is_oopptr()->instance_id();
520 int new_idx = new_index(old_idx);
521 const Type* new_type = type->is_oopptr()->with_instance_id(new_idx);
522 _new_type_array.map(n->_idx, new_type);
523 no_of_updates++;
524 }
525
526 return no_of_updates;
527 }
528
529 void PhaseValues::init_con_caches() {
530 memset(_icons,0,sizeof(_icons));
531 memset(_lcons,0,sizeof(_lcons));
532 memset(_zcons,0,sizeof(_zcons));
533 }
534
535 //--------------------------------find_int_type--------------------------------
536 const TypeInt* PhaseValues::find_int_type(Node* n) {
537 if (n == nullptr) return nullptr;
538 // Call type_or_null(n) to determine node's type since we might be in
539 // parse phase and call n->Value() may return wrong type.
540 // (For example, a phi node at the beginning of loop parsing is not ready.)
541 const Type* t = type_or_null(n);
542 if (t == nullptr) return nullptr;
543 return t->isa_int();
544 }
545
546
547 //-------------------------------find_long_type--------------------------------
548 const TypeLong* PhaseValues::find_long_type(Node* n) {
549 if (n == nullptr) return nullptr;
550 // (See comment above on type_or_null.)
551 const Type* t = type_or_null(n);
552 if (t == nullptr) return nullptr;
553 return t->isa_long();
554 }
555
556 //------------------------------~PhaseValues-----------------------------------
557 #ifndef PRODUCT
558 PhaseValues::~PhaseValues() {
559 // Statistics for NodeHash
560 _table.dump();
561 // Statistics for value progress and efficiency
562 if( PrintCompilation && Verbose && WizardMode ) {
563 tty->print("\n%sValues: %d nodes ---> %d/%d (%d)",
564 is_IterGVN() ? "Iter" : " ", C->unique(), made_progress(), made_transforms(), made_new_values());
565 if( made_transforms() != 0 ) {
566 tty->print_cr(" ratio %f", made_progress()/(float)made_transforms() );
567 } else {
568 tty->cr();
569 }
570 }
571 }
572 #endif
573
574 //------------------------------makecon----------------------------------------
575 ConNode* PhaseValues::makecon(const Type* t) {
576 assert(t->singleton(), "must be a constant");
577 assert(!t->empty() || t == Type::TOP, "must not be vacuous range");
578 switch (t->base()) { // fast paths
579 case Type::Half:
580 case Type::Top: return (ConNode*) C->top();
581 case Type::Int: return intcon( t->is_int()->get_con() );
582 case Type::Long: return longcon( t->is_long()->get_con() );
583 default: break;
584 }
585 if (t->is_zero_type())
586 return zerocon(t->basic_type());
587 return uncached_makecon(t);
588 }
589
590 //--------------------------uncached_makecon-----------------------------------
591 // Make an idealized constant - one of ConINode, ConPNode, etc.
592 ConNode* PhaseValues::uncached_makecon(const Type *t) {
593 assert(t->singleton(), "must be a constant");
594 ConNode* x = ConNode::make(t);
595 ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering
596 if (k == nullptr) {
597 set_type(x, t); // Missed, provide type mapping
598 GrowableArray<Node_Notes*>* nna = C->node_note_array();
599 if (nna != nullptr) {
600 Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true);
601 loc->clear(); // do not put debug info on constants
602 }
603 } else {
604 x->destruct(this); // Hit, destroy duplicate constant
605 x = k; // use existing constant
606 }
607 return x;
608 }
609
610 //------------------------------intcon-----------------------------------------
611 // Fast integer constant. Same as "transform(new ConINode(TypeInt::make(i)))"
612 ConINode* PhaseValues::intcon(jint i) {
613 // Small integer? Check cache! Check that cached node is not dead
614 if (i >= _icon_min && i <= _icon_max) {
615 ConINode* icon = _icons[i-_icon_min];
616 if (icon != nullptr && icon->in(TypeFunc::Control) != nullptr)
617 return icon;
618 }
619 ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i));
620 assert(icon->is_Con(), "");
621 if (i >= _icon_min && i <= _icon_max)
622 _icons[i-_icon_min] = icon; // Cache small integers
623 return icon;
624 }
625
626 //------------------------------longcon----------------------------------------
627 // Fast long constant.
628 ConLNode* PhaseValues::longcon(jlong l) {
629 // Small integer? Check cache! Check that cached node is not dead
630 if (l >= _lcon_min && l <= _lcon_max) {
631 ConLNode* lcon = _lcons[l-_lcon_min];
632 if (lcon != nullptr && lcon->in(TypeFunc::Control) != nullptr)
633 return lcon;
634 }
635 ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l));
636 assert(lcon->is_Con(), "");
637 if (l >= _lcon_min && l <= _lcon_max)
638 _lcons[l-_lcon_min] = lcon; // Cache small integers
639 return lcon;
640 }
641 ConNode* PhaseValues::integercon(jlong l, BasicType bt) {
642 if (bt == T_INT) {
643 return intcon(checked_cast<jint>(l));
644 }
645 assert(bt == T_LONG, "not an integer");
646 return longcon(l);
647 }
648
649
650 //------------------------------zerocon-----------------------------------------
651 // Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))"
652 ConNode* PhaseValues::zerocon(BasicType bt) {
653 assert((uint)bt <= _zcon_max, "domain check");
654 ConNode* zcon = _zcons[bt];
655 if (zcon != nullptr && zcon->in(TypeFunc::Control) != nullptr)
656 return zcon;
657 zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt));
658 _zcons[bt] = zcon;
659 return zcon;
660 }
661
662
663
664 //=============================================================================
665 Node* PhaseGVN::apply_ideal(Node* k, bool can_reshape) {
666 Node* i = BarrierSet::barrier_set()->barrier_set_c2()->ideal_node(this, k, can_reshape);
667 if (i == nullptr) {
668 i = k->Ideal(this, can_reshape);
669 }
670 return i;
671 }
672
673 //------------------------------transform--------------------------------------
674 // Return a node which computes the same function as this node, but
675 // in a faster or cheaper fashion.
676 Node* PhaseGVN::transform(Node* n) {
677 NOT_PRODUCT( set_transforms(); )
678
679 // Apply the Ideal call in a loop until it no longer applies
680 Node* k = n;
681 Node* i = apply_ideal(k, /*can_reshape=*/false);
682 NOT_PRODUCT(uint loop_count = 1;)
683 while (i != nullptr) {
684 assert(i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" );
685 k = i;
686 #ifdef ASSERT
687 if (loop_count >= K + C->live_nodes()) {
688 dump_infinite_loop_info(i, "PhaseGVN::transform");
689 }
690 #endif
691 i = apply_ideal(k, /*can_reshape=*/false);
692 NOT_PRODUCT(loop_count++;)
693 }
694 NOT_PRODUCT(if (loop_count != 0) { set_progress(); })
695
696 // If brand new node, make space in type array.
697 ensure_type_or_null(k);
698
699 // Since I just called 'Value' to compute the set of run-time values
700 // for this Node, and 'Value' is non-local (and therefore expensive) I'll
701 // cache Value. Later requests for the local phase->type of this Node can
702 // use the cached Value instead of suffering with 'bottom_type'.
703 const Type* t = k->Value(this); // Get runtime Value set
704 assert(t != nullptr, "value sanity");
705 if (type_or_null(k) != t) {
706 #ifndef PRODUCT
707 // Do not count initial visit to node as a transformation
708 if (type_or_null(k) == nullptr) {
709 inc_new_values();
710 set_progress();
711 }
712 #endif
713 set_type(k, t);
714 // If k is a TypeNode, capture any more-precise type permanently into Node
715 k->raise_bottom_type(t);
716 }
717
718 if (t->singleton() && !k->is_Con()) {
719 NOT_PRODUCT(set_progress();)
720 return makecon(t); // Turn into a constant
721 }
722
723 // Now check for Identities
724 i = k->Identity(this); // Look for a nearby replacement
725 if (i != k) { // Found? Return replacement!
726 NOT_PRODUCT(set_progress();)
727 return i;
728 }
729
730 // Global Value Numbering
731 i = hash_find_insert(k); // Insert if new
732 if (i && (i != k)) {
733 // Return the pre-existing node
734 NOT_PRODUCT(set_progress();)
735 return i;
736 }
737
738 // Return Idealized original
739 return k;
740 }
741
742 bool PhaseGVN::is_dominator_helper(Node *d, Node *n, bool linear_only) {
743 if (d->is_top() || (d->is_Proj() && d->in(0)->is_top())) {
744 return false;
745 }
746 if (n->is_top() || (n->is_Proj() && n->in(0)->is_top())) {
747 return false;
748 }
749 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
750 int i = 0;
751 while (d != n) {
752 n = IfNode::up_one_dom(n, linear_only);
753 i++;
754 if (n == nullptr || i >= 100) {
755 return false;
756 }
757 }
758 return true;
759 }
760
761 #ifdef ASSERT
762 //------------------------------dead_loop_check--------------------------------
763 // Check for a simple dead loop when a data node references itself directly
764 // or through an other data node excluding cons and phis.
765 void PhaseGVN::dead_loop_check( Node *n ) {
766 // Phi may reference itself in a loop
767 if (n != nullptr && !n->is_dead_loop_safe() && !n->is_CFG()) {
768 // Do 2 levels check and only data inputs.
769 bool no_dead_loop = true;
770 uint cnt = n->req();
771 for (uint i = 1; i < cnt && no_dead_loop; i++) {
772 Node *in = n->in(i);
773 if (in == n) {
774 no_dead_loop = false;
775 } else if (in != nullptr && !in->is_dead_loop_safe()) {
776 uint icnt = in->req();
777 for (uint j = 1; j < icnt && no_dead_loop; j++) {
778 if (in->in(j) == n || in->in(j) == in)
779 no_dead_loop = false;
780 }
781 }
782 }
783 if (!no_dead_loop) n->dump_bfs(100,nullptr,"#");
784 assert(no_dead_loop, "dead loop detected");
785 }
786 }
787
788
789 /**
790 * Dumps information that can help to debug the problem. A debug
791 * build fails with an assert.
792 */
793 void PhaseGVN::dump_infinite_loop_info(Node* n, const char* where) {
794 n->dump(4);
795 assert(false, "infinite loop in %s", where);
796 }
797 #endif
798
799 //=============================================================================
800 //------------------------------PhaseIterGVN-----------------------------------
801 // Initialize with previous PhaseIterGVN info; used by PhaseCCP
802 PhaseIterGVN::PhaseIterGVN(PhaseIterGVN* igvn) : _delay_transform(igvn->_delay_transform),
803 _worklist(*C->igvn_worklist())
804 {
805 _iterGVN = true;
806 assert(&_worklist == &igvn->_worklist, "sanity");
807 }
808
809 //------------------------------PhaseIterGVN-----------------------------------
810 // Initialize with previous PhaseGVN info from Parser
811 PhaseIterGVN::PhaseIterGVN(PhaseGVN* gvn) : _delay_transform(false),
812 _worklist(*C->igvn_worklist())
813 {
814 _iterGVN = true;
815 uint max;
816
817 // Dead nodes in the hash table inherited from GVN were not treated as
818 // roots during def-use info creation; hence they represent an invisible
819 // use. Clear them out.
820 max = _table.size();
821 for( uint i = 0; i < max; ++i ) {
822 Node *n = _table.at(i);
823 if(n != nullptr && n != _table.sentinel() && n->outcnt() == 0) {
824 if( n->is_top() ) continue;
825 // If remove_useless_nodes() has run, we expect no such nodes left.
826 assert(false, "remove_useless_nodes missed this node");
827 hash_delete(n);
828 }
829 }
830
831 // Any Phis or Regions on the worklist probably had uses that could not
832 // make more progress because the uses were made while the Phis and Regions
833 // were in half-built states. Put all uses of Phis and Regions on worklist.
834 max = _worklist.size();
835 for( uint j = 0; j < max; j++ ) {
836 Node *n = _worklist.at(j);
837 uint uop = n->Opcode();
838 if( uop == Op_Phi || uop == Op_Region ||
839 n->is_Type() ||
840 n->is_Mem() )
841 add_users_to_worklist(n);
842 }
843 }
844
845 void PhaseIterGVN::shuffle_worklist() {
846 if (_worklist.size() < 2) return;
847 for (uint i = _worklist.size() - 1; i >= 1; i--) {
848 uint j = C->random() % (i + 1);
849 swap(_worklist.adr()[i], _worklist.adr()[j]);
850 }
851 }
852
853 #ifndef PRODUCT
854 void PhaseIterGVN::verify_step(Node* n) {
855 if (is_verify_def_use()) {
856 ResourceMark rm;
857 VectorSet visited;
858 Node_List worklist;
859
860 _verify_window[_verify_counter % _verify_window_size] = n;
861 ++_verify_counter;
862 if (C->unique() < 1000 || 0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) {
863 ++_verify_full_passes;
864 worklist.push(C->root());
865 Node::verify(-1, visited, worklist);
866 return;
867 }
868 for (int i = 0; i < _verify_window_size; i++) {
869 Node* n = _verify_window[i];
870 if (n == nullptr) {
871 continue;
872 }
873 if (n->in(0) == NodeSentinel) { // xform_idom
874 _verify_window[i] = n->in(1);
875 --i;
876 continue;
877 }
878 // Typical fanout is 1-2, so this call visits about 6 nodes.
879 if (!visited.test_set(n->_idx)) {
880 worklist.push(n);
881 }
882 }
883 Node::verify(4, visited, worklist);
884 }
885 }
886
887 void PhaseIterGVN::trace_PhaseIterGVN(Node* n, Node* nn, const Type* oldtype) {
888 const Type* newtype = type_or_null(n);
889 if (nn != n || oldtype != newtype) {
890 C->print_method(PHASE_AFTER_ITER_GVN_STEP, 5, n);
891 }
892 if (TraceIterativeGVN) {
893 uint wlsize = _worklist.size();
894 if (nn != n) {
895 // print old node
896 tty->print("< ");
897 if (oldtype != newtype && oldtype != nullptr) {
898 oldtype->dump();
899 }
900 do { tty->print("\t"); } while (tty->position() < 16);
901 tty->print("<");
902 n->dump();
903 }
904 if (oldtype != newtype || nn != n) {
905 // print new node and/or new type
906 if (oldtype == nullptr) {
907 tty->print("* ");
908 } else if (nn != n) {
909 tty->print("> ");
910 } else {
911 tty->print("= ");
912 }
913 if (newtype == nullptr) {
914 tty->print("null");
915 } else {
916 newtype->dump();
917 }
918 do { tty->print("\t"); } while (tty->position() < 16);
919 nn->dump();
920 }
921 if (Verbose && wlsize < _worklist.size()) {
922 tty->print(" Push {");
923 while (wlsize != _worklist.size()) {
924 Node* pushed = _worklist.at(wlsize++);
925 tty->print(" %d", pushed->_idx);
926 }
927 tty->print_cr(" }");
928 }
929 if (nn != n) {
930 // ignore n, it might be subsumed
931 verify_step((Node*) nullptr);
932 }
933 }
934 }
935
936 void PhaseIterGVN::init_verifyPhaseIterGVN() {
937 _verify_counter = 0;
938 _verify_full_passes = 0;
939 for (int i = 0; i < _verify_window_size; i++) {
940 _verify_window[i] = nullptr;
941 }
942 #ifdef ASSERT
943 // Verify that all modified nodes are on _worklist
944 Unique_Node_List* modified_list = C->modified_nodes();
945 while (modified_list != nullptr && modified_list->size()) {
946 Node* n = modified_list->pop();
947 if (!n->is_Con() && !_worklist.member(n)) {
948 n->dump();
949 fatal("modified node is not on IGVN._worklist");
950 }
951 }
952 #endif
953 }
954
955 void PhaseIterGVN::verify_PhaseIterGVN() {
956 #ifdef ASSERT
957 // Verify nodes with changed inputs.
958 Unique_Node_List* modified_list = C->modified_nodes();
959 while (modified_list != nullptr && modified_list->size()) {
960 Node* n = modified_list->pop();
961 if (!n->is_Con()) { // skip Con nodes
962 n->dump();
963 fatal("modified node was not processed by IGVN.transform_old()");
964 }
965 }
966 #endif
967
968 C->verify_graph_edges();
969 if (is_verify_def_use() && PrintOpto) {
970 if (_verify_counter == _verify_full_passes) {
971 tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes",
972 (int) _verify_full_passes);
973 } else {
974 tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes",
975 (int) _verify_counter, (int) _verify_full_passes);
976 }
977 }
978
979 #ifdef ASSERT
980 if (modified_list != nullptr) {
981 while (modified_list->size() > 0) {
982 Node* n = modified_list->pop();
983 n->dump();
984 assert(false, "VerifyIterativeGVN: new modified node was added");
985 }
986 }
987
988 verify_optimize();
989 #endif
990 }
991 #endif /* PRODUCT */
992
993 #ifdef ASSERT
994 /**
995 * Dumps information that can help to debug the problem. A debug
996 * build fails with an assert.
997 */
998 void PhaseIterGVN::dump_infinite_loop_info(Node* n, const char* where) {
999 n->dump(4);
1000 _worklist.dump();
1001 assert(false, "infinite loop in %s", where);
1002 }
1003
1004 /**
1005 * Prints out information about IGVN if the 'verbose' option is used.
1006 */
1007 void PhaseIterGVN::trace_PhaseIterGVN_verbose(Node* n, int num_processed) {
1008 if (TraceIterativeGVN && Verbose) {
1009 tty->print(" Pop ");
1010 n->dump();
1011 if ((num_processed % 100) == 0) {
1012 _worklist.print_set();
1013 }
1014 }
1015 }
1016 #endif /* ASSERT */
1017
1018 void PhaseIterGVN::optimize() {
1019 DEBUG_ONLY(uint num_processed = 0;)
1020 NOT_PRODUCT(init_verifyPhaseIterGVN();)
1021 NOT_PRODUCT(C->reset_igv_phase_iter(PHASE_AFTER_ITER_GVN_STEP);)
1022 C->print_method(PHASE_BEFORE_ITER_GVN, 3);
1023 if (StressIGVN) {
1024 shuffle_worklist();
1025 }
1026
1027 // The node count check in the loop below (check_node_count) assumes that we
1028 // increase the live node count with at most
1029 // max_live_nodes_increase_per_iteration in between checks. If this
1030 // assumption does not hold, there is a risk that we exceed the max node
1031 // limit in between checks and trigger an assert during node creation.
1032 const int max_live_nodes_increase_per_iteration = NodeLimitFudgeFactor * 3;
1033
1034 uint loop_count = 0;
1035 // Pull from worklist and transform the node. If the node has changed,
1036 // update edge info and put uses on worklist.
1037 while (_worklist.size() > 0) {
1038 if (C->check_node_count(max_live_nodes_increase_per_iteration, "Out of nodes")) {
1039 C->print_method(PHASE_AFTER_ITER_GVN, 3);
1040 return;
1041 }
1042 Node* n = _worklist.pop();
1043 if (loop_count >= K * C->live_nodes()) {
1044 DEBUG_ONLY(dump_infinite_loop_info(n, "PhaseIterGVN::optimize");)
1045 C->record_method_not_compilable("infinite loop in PhaseIterGVN::optimize");
1046 C->print_method(PHASE_AFTER_ITER_GVN, 3);
1047 return;
1048 }
1049 DEBUG_ONLY(trace_PhaseIterGVN_verbose(n, num_processed++);)
1050 if (n->outcnt() != 0) {
1051 NOT_PRODUCT(const Type* oldtype = type_or_null(n));
1052 // Do the transformation
1053 DEBUG_ONLY(int live_nodes_before = C->live_nodes();)
1054 Node* nn = transform_old(n);
1055 DEBUG_ONLY(int live_nodes_after = C->live_nodes();)
1056 // Ensure we did not increase the live node count with more than
1057 // max_live_nodes_increase_per_iteration during the call to transform_old
1058 DEBUG_ONLY(int increase = live_nodes_after - live_nodes_before;)
1059 assert(increase < max_live_nodes_increase_per_iteration,
1060 "excessive live node increase in single iteration of IGVN: %d "
1061 "(should be at most %d)",
1062 increase, max_live_nodes_increase_per_iteration);
1063 NOT_PRODUCT(trace_PhaseIterGVN(n, nn, oldtype);)
1064 } else if (!n->is_top()) {
1065 remove_dead_node(n);
1066 }
1067 loop_count++;
1068 }
1069 NOT_PRODUCT(verify_PhaseIterGVN();)
1070 C->print_method(PHASE_AFTER_ITER_GVN, 3);
1071 }
1072
1073 #ifdef ASSERT
1074 void PhaseIterGVN::verify_optimize() {
1075 assert(_worklist.size() == 0, "igvn worklist must be empty before verify");
1076
1077 if (is_verify_Value() ||
1078 is_verify_Ideal() ||
1079 is_verify_Identity()) {
1080 ResourceMark rm;
1081 Unique_Node_List worklist;
1082 bool failure = false;
1083 // BFS all nodes, starting at root
1084 worklist.push(C->root());
1085 for (uint j = 0; j < worklist.size(); ++j) {
1086 Node* n = worklist.at(j);
1087 if (is_verify_Value()) { failure |= verify_Value_for(n); }
1088 if (is_verify_Ideal()) { failure |= verify_Ideal_for(n, false); }
1089 if (is_verify_Ideal()) { failure |= verify_Ideal_for(n, true); }
1090 if (is_verify_Identity()) { failure |= verify_Identity_for(n); }
1091 // traverse all inputs and outputs
1092 for (uint i = 0; i < n->req(); i++) {
1093 if (n->in(i) != nullptr) {
1094 worklist.push(n->in(i));
1095 }
1096 }
1097 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1098 worklist.push(n->fast_out(i));
1099 }
1100 }
1101 // If we get this assert, check why the reported nodes were not processed again in IGVN.
1102 // We should either make sure that these nodes are properly added back to the IGVN worklist
1103 // in PhaseIterGVN::add_users_to_worklist to update them again or add an exception
1104 // in the verification code above if that is not possible for some reason (like Load nodes).
1105 assert(!failure, "Missed optimization opportunity in PhaseIterGVN");
1106 }
1107
1108 verify_empty_worklist(nullptr);
1109 }
1110
1111 void PhaseIterGVN::verify_empty_worklist(Node* node) {
1112 // Verify that the igvn worklist is empty. If no optimization happened, then
1113 // nothing needs to be on the worklist.
1114 if (_worklist.size() == 0) { return; }
1115
1116 stringStream ss; // Print as a block without tty lock.
1117 for (uint j = 0; j < _worklist.size(); j++) {
1118 Node* n = _worklist.at(j);
1119 ss.print("igvn.worklist[%d] ", j);
1120 n->dump("\n", false, &ss);
1121 }
1122 if (_worklist.size() != 0 && node != nullptr) {
1123 ss.print_cr("Previously optimized:");
1124 node->dump("\n", false, &ss);
1125 }
1126 tty->print_cr("%s", ss.as_string());
1127 assert(false, "igvn worklist must still be empty after verify");
1128 }
1129
1130 // Check that type(n) == n->Value(), return true if we have a failure.
1131 // We have a list of exceptions, see detailed comments in code.
1132 // (1) Integer "widen" changes, but the range is the same.
1133 // (2) LoadNode performs deep traversals. Load is not notified for changes far away.
1134 // (3) CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
1135 bool PhaseIterGVN::verify_Value_for(Node* n) {
1136 // If we assert inside type(n), because the type is still a null, then maybe
1137 // the node never went through gvn.transform, which would be a bug.
1138 const Type* told = type(n);
1139 const Type* tnew = n->Value(this);
1140 if (told == tnew) {
1141 return false;
1142 }
1143 // Exception (1)
1144 // Integer "widen" changes, but range is the same.
1145 if (told->isa_integer(tnew->basic_type()) != nullptr) { // both either int or long
1146 const TypeInteger* t0 = told->is_integer(tnew->basic_type());
1147 const TypeInteger* t1 = tnew->is_integer(tnew->basic_type());
1148 if (t0->lo_as_long() == t1->lo_as_long() &&
1149 t0->hi_as_long() == t1->hi_as_long()) {
1150 return false; // ignore integer widen
1151 }
1152 }
1153 // Exception (2)
1154 // LoadNode performs deep traversals. Load is not notified for changes far away.
1155 if (n->is_Load() && !told->singleton()) {
1156 // MemNode::can_see_stored_value looks up through many memory nodes,
1157 // which means we would need to notify modifications from far up in
1158 // the inputs all the way down to the LoadNode. We don't do that.
1159 return false;
1160 }
1161 // Exception (3)
1162 // CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away.
1163 if (n->Opcode() == Op_CmpP && type(n->in(1))->isa_oopptr() && type(n->in(2))->isa_oopptr()) {
1164 // SubNode::Value
1165 // CmpPNode::sub
1166 // MemNode::detect_ptr_independence
1167 // MemNode::all_controls_dominate
1168 // We find all controls of a pointer load, and see if they dominate the control of
1169 // an allocation. If they all dominate, we know the allocation is after (independent)
1170 // of the pointer load, and we can say the pointers are different. For this we call
1171 // n->dominates(sub, nlist) to check if controls n of the pointer load dominate the
1172 // control sub of the allocation. The problems is that sometimes dominates answers
1173 // false conservatively, and later it can determine that it is indeed true. Loops with
1174 // Region heads can lead to giving up, whereas LoopNodes can be skipped easier, and
1175 // so the traversal becomes more powerful. This is difficult to remidy, we would have
1176 // to notify the CmpP of CFG updates. Luckily, we recompute CmpP::Value during CCP
1177 // after loop-opts, so that should take care of many of these cases.
1178 return false;
1179 }
1180
1181 stringStream ss; // Print as a block without tty lock.
1182 ss.cr();
1183 ss.print_cr("Missed Value optimization:");
1184 n->dump_bfs(3, nullptr, "", &ss);
1185 ss.print_cr("Current type:");
1186 told->dump_on(&ss);
1187 ss.cr();
1188 ss.print_cr("Optimized type:");
1189 tnew->dump_on(&ss);
1190 ss.cr();
1191 tty->print_cr("%s", ss.as_string());
1192 return true;
1193 }
1194
1195 // Check that all Ideal optimizations that could be done were done.
1196 // Returns true if it found missed optimization opportunities and
1197 // false otherwise (no missed optimization, or skipped verification).
1198 bool PhaseIterGVN::verify_Ideal_for(Node* n, bool can_reshape) {
1199 // First, we check a list of exceptions, where we skip verification,
1200 // because there are known cases where Ideal can optimize after IGVN.
1201 // Some may be expected and cannot be fixed, and others should be fixed.
1202 switch (n->Opcode()) {
1203 // RangeCheckNode::Ideal looks up the chain for about 999 nodes
1204 // (see "Range-Check scan limit"). So, it is possible that something
1205 // is optimized in that input subgraph, and the RangeCheck was not
1206 // added to the worklist because it would be too expensive to walk
1207 // down the graph for 1000 nodes and put all on the worklist.
1208 //
1209 // Found with:
1210 // java -XX:VerifyIterativeGVN=0100 -Xbatch --version
1211 case Op_RangeCheck:
1212 return false;
1213
1214 // IfNode::Ideal does:
1215 // Node* prev_dom = search_identical(dist, igvn);
1216 // which means we seach up the CFG, traversing at most up to a distance.
1217 // If anything happens rather far away from the If, we may not put the If
1218 // back on the worklist.
1219 //
1220 // Found with:
1221 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1222 case Op_If:
1223 return false;
1224
1225 // IfNode::simple_subsuming
1226 // Looks for dominating test that subsumes the current test.
1227 // Notification could be difficult because of larger distance.
1228 //
1229 // Found with:
1230 // runtime/exceptionMsgs/ArrayIndexOutOfBoundsException/ArrayIndexOutOfBoundsExceptionTest.java#id1
1231 // -XX:VerifyIterativeGVN=1110
1232 case Op_CountedLoopEnd:
1233 return false;
1234
1235 // LongCountedLoopEndNode::Ideal
1236 // Probably same issue as above.
1237 //
1238 // Found with:
1239 // compiler/predicates/assertion/TestAssertionPredicates.java#NoLoopPredicationXbatch
1240 // -XX:StressLongCountedLoop=2000000 -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1241 case Op_LongCountedLoopEnd:
1242 return false;
1243
1244 // RegionNode::Ideal does "Skip around the useless IF diamond".
1245 // 245 IfTrue === 244
1246 // 258 If === 245 257
1247 // 259 IfTrue === 258 [[ 263 ]]
1248 // 260 IfFalse === 258 [[ 263 ]]
1249 // 263 Region === 263 260 259 [[ 263 268 ]]
1250 // to
1251 // 245 IfTrue === 244
1252 // 263 Region === 263 245 _ [[ 263 268 ]]
1253 //
1254 // "Useless" means that there is no code in either branch of the If.
1255 // I found a case where this was not done yet during IGVN.
1256 // Why does the Region not get added to IGVN worklist when the If diamond becomes useless?
1257 //
1258 // Found with:
1259 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1260 case Op_Region:
1261 return false;
1262
1263 // In AddNode::Ideal, we call "commute", which swaps the inputs so
1264 // that smaller idx are first. Tracking it back, it led me to
1265 // PhaseIdealLoop::remix_address_expressions which swapped the edges.
1266 //
1267 // Example:
1268 // Before PhaseIdealLoop::remix_address_expressions
1269 // 154 AddI === _ 12 144
1270 // After PhaseIdealLoop::remix_address_expressions
1271 // 154 AddI === _ 144 12
1272 // After AddNode::Ideal
1273 // 154 AddI === _ 12 144
1274 //
1275 // I suspect that the node should be added to the IGVN worklist after
1276 // PhaseIdealLoop::remix_address_expressions
1277 //
1278 // This is the only case I looked at, there may be others. Found like this:
1279 // java -XX:VerifyIterativeGVN=0100 -Xbatch --version
1280 //
1281 // The following hit the same logic in PhaseIdealLoop::remix_address_expressions.
1282 //
1283 // Note: currently all of these fail also for other reasons, for example
1284 // because of "commute" doing the reordering with the phi below. Once
1285 // that is resolved, we can come back to this issue here.
1286 //
1287 // case Op_AddD:
1288 // case Op_AddI:
1289 // case Op_AddL:
1290 // case Op_AddF:
1291 // case Op_MulI:
1292 // case Op_MulL:
1293 // case Op_MulF:
1294 // case Op_MulD:
1295 // if (n->in(1)->_idx > n->in(2)->_idx) {
1296 // // Expect "commute" to revert this case.
1297 // return false;
1298 // }
1299 // break; // keep verifying
1300
1301 // AddFNode::Ideal calls "commute", which can reorder the inputs for this:
1302 // Check for tight loop increments: Loop-phi of Add of loop-phi
1303 // It wants to take the phi into in(1):
1304 // 471 Phi === 435 38 390
1305 // 390 AddF === _ 471 391
1306 //
1307 // Other Associative operators are also affected equally.
1308 //
1309 // Investigate why this does not happen earlier during IGVN.
1310 //
1311 // Found with:
1312 // test/hotspot/jtreg/compiler/loopopts/superword/ReductionPerf.java
1313 // -XX:VerifyIterativeGVN=1110
1314 case Op_AddD:
1315 //case Op_AddI: // Also affected for other reasons, see case further down.
1316 //case Op_AddL: // Also affected for other reasons, see case further down.
1317 case Op_AddF:
1318 case Op_MulI:
1319 case Op_MulL:
1320 case Op_MulF:
1321 case Op_MulD:
1322 case Op_MinF:
1323 case Op_MinD:
1324 case Op_MaxF:
1325 case Op_MaxD:
1326 // XorINode::Ideal
1327 // Found with:
1328 // compiler/intrinsics/chacha/TestChaCha20.java
1329 // -XX:VerifyIterativeGVN=1110
1330 case Op_XorI:
1331 case Op_XorL:
1332 // It seems we may have similar issues with the HF cases.
1333 // Found with aarch64:
1334 // compiler/vectorization/TestFloat16VectorOperations.java
1335 // -XX:VerifyIterativeGVN=1110
1336 case Op_AddHF:
1337 case Op_MulHF:
1338 case Op_MaxHF:
1339 case Op_MinHF:
1340 return false;
1341
1342 // In MulNode::Ideal the edges can be swapped to help value numbering:
1343 //
1344 // // We are OK if right is a constant, or right is a load and
1345 // // left is a non-constant.
1346 // if( !(t2->singleton() ||
1347 // (in(2)->is_Load() && !(t1->singleton() || in(1)->is_Load())) ) ) {
1348 // if( t1->singleton() || // Left input is a constant?
1349 // // Otherwise, sort inputs (commutativity) to help value numbering.
1350 // (in(1)->_idx > in(2)->_idx) ) {
1351 // swap_edges(1, 2);
1352 //
1353 // Why was this not done earlier during IGVN?
1354 //
1355 // Found with:
1356 // test/hotspot/jtreg/gc/stress/gcbasher/TestGCBasherWithG1.java
1357 // -XX:VerifyIterativeGVN=1110
1358 case Op_AndI:
1359 // Same for AndL.
1360 // Found with:
1361 // compiler/intrinsics/bigInteger/MontgomeryMultiplyTest.java
1362 // -XX:VerifyIterativeGVN=1110
1363 case Op_AndL:
1364 return false;
1365
1366 // SubLNode::Ideal does transform like:
1367 // Convert "c1 - (y+c0)" into "(c1-c0) - y"
1368 //
1369 // In IGVN before verification:
1370 // 8423 ConvI2L === _ 3519 [[ 8424 ]] #long:-2
1371 // 8422 ConvI2L === _ 8399 [[ 8424 ]] #long:3..256:www
1372 // 8424 AddL === _ 8422 8423 [[ 8383 ]] !orig=[8382]
1373 // 8016 ConL === 0 [[ 8383 ]] #long:0
1374 // 8383 SubL === _ 8016 8424 [[ 8156 ]] !orig=[8154]
1375 //
1376 // And then in verification:
1377 // 8338 ConL === 0 [[ 8339 8424 ]] #long:-2 <----- Was constant folded.
1378 // 8422 ConvI2L === _ 8399 [[ 8424 ]] #long:3..256:www
1379 // 8424 AddL === _ 8422 8338 [[ 8383 ]] !orig=[8382]
1380 // 8016 ConL === 0 [[ 8383 ]] #long:0
1381 // 8383 SubL === _ 8016 8424 [[ 8156 ]] !orig=[8154]
1382 //
1383 // So the form changed from:
1384 // c1 - (y + [8423 ConvI2L])
1385 // to
1386 // c1 - (y + -2)
1387 // but the SubL was not added to the IGVN worklist. Investigate why.
1388 // There could be other issues too.
1389 //
1390 // There seems to be a related AddL IGVN optimization that triggers
1391 // the same SubL optimization, so investigate that too.
1392 //
1393 // Found with:
1394 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1395 case Op_SubL:
1396 return false;
1397
1398 // SubINode::Ideal does
1399 // Convert "x - (y+c0)" into "(x-y) - c0" AND
1400 // Convert "c1 - (y+c0)" into "(c1-c0) - y"
1401 //
1402 // Investigate why this does not yet happen during IGVN.
1403 //
1404 // Found with:
1405 // test/hotspot/jtreg/compiler/c2/IVTest.java
1406 // -XX:VerifyIterativeGVN=1110
1407 case Op_SubI:
1408 return false;
1409
1410 // AddNode::IdealIL does transform like:
1411 // Convert x + (con - y) into "(x - y) + con"
1412 //
1413 // In IGVN before verification:
1414 // 8382 ConvI2L
1415 // 8381 ConvI2L === _ 791 [[ 8383 ]] #long:0
1416 // 8383 SubL === _ 8381 8382
1417 // 8168 ConvI2L
1418 // 8156 AddL === _ 8168 8383 [[ 8158 ]]
1419 //
1420 // And then in verification:
1421 // 8424 AddL
1422 // 8016 ConL === 0 [[ 8383 ]] #long:0 <--- Was constant folded.
1423 // 8383 SubL === _ 8016 8424
1424 // 8168 ConvI2L
1425 // 8156 AddL === _ 8168 8383 [[ 8158 ]]
1426 //
1427 // So the form changed from:
1428 // x + (ConvI2L(0) - [8382 ConvI2L])
1429 // to
1430 // x + (0 - [8424 AddL])
1431 // but the AddL was not added to the IGVN worklist. Investigate why.
1432 // There could be other issues, too. For example with "commute", see above.
1433 //
1434 // Found with:
1435 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1436 case Op_AddL:
1437 return false;
1438
1439 // SubTypeCheckNode::Ideal calls SubTypeCheckNode::verify_helper, which does
1440 // Node* cmp = phase->transform(new CmpPNode(subklass, in(SuperKlass)));
1441 // record_for_cleanup(cmp, phase);
1442 // This verification code in the Ideal code creates new nodes, and checks
1443 // if they fold in unexpected ways. This means some nodes are created and
1444 // added to the worklist, even if the SubTypeCheck is not optimized. This
1445 // goes agains the assumption of the verification here, which assumes that
1446 // if the node is not optimized, then no new nodes should be created, and
1447 // also no nodes should be added to the worklist.
1448 // I see two options:
1449 // 1) forbid what verify_helper does, because for each Ideal call it
1450 // uses memory and that is suboptimal. But it is not clear how that
1451 // verification can be done otherwise.
1452 // 2) Special case the verification here. Probably the new nodes that
1453 // were just created are dead, i.e. they are not connected down to
1454 // root. We could verify that, and remove those nodes from the graph
1455 // by setting all their inputs to nullptr. And of course we would
1456 // have to remove those nodes from the worklist.
1457 // Maybe there are other options too, I did not dig much deeper yet.
1458 //
1459 // Found with:
1460 // java -XX:VerifyIterativeGVN=0100 -Xbatch --version
1461 case Op_SubTypeCheck:
1462 return false;
1463
1464 // LoopLimitNode::Ideal when stride is constant power-of-2, we can do a lowering
1465 // to other nodes: Conv, Add, Sub, Mul, And ...
1466 //
1467 // 107 ConI === 0 [[ ... ]] #int:2
1468 // 84 LoadRange === _ 7 83
1469 // 50 ConI === 0 [[ ... ]] #int:0
1470 // 549 LoopLimit === _ 50 84 107
1471 //
1472 // I stepped backward, to see how the node was generated, and I found that it was
1473 // created in PhaseIdealLoop::exact_limit and not changed since. It is added to the
1474 // IGVN worklist. I quickly checked when it goes into LoopLimitNode::Ideal after
1475 // that, and it seems we want to skip lowering it until after loop-opts, but never
1476 // add call record_for_post_loop_opts_igvn. This would be an easy fix, but there
1477 // could be other issues too.
1478 //
1479 // Fond with:
1480 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1481 case Op_LoopLimit:
1482 return false;
1483
1484 // PhiNode::Ideal calls split_flow_path, which tries to do this:
1485 // "This optimization tries to find two or more inputs of phi with the same constant
1486 // value. It then splits them into a separate Phi, and according Region."
1487 //
1488 // Example:
1489 // 130 DecodeN === _ 129
1490 // 50 ConP === 0 [[ 18 91 99 18 ]] #null
1491 // 18 Phi === 14 50 130 50 [[ 133 ]] #java/lang/Object * Oop:java/lang/Object *
1492 //
1493 // turns into:
1494 //
1495 // 50 ConP === 0 [[ 99 91 18 ]] #null
1496 // 130 DecodeN === _ 129 [[ 18 ]]
1497 // 18 Phi === 14 130 50 [[ 133 ]] #java/lang/Object * Oop:java/lang/Object *
1498 //
1499 // We would have to investigate why this optimization does not happen during IGVN.
1500 // There could also be other issues - I did not investigate further yet.
1501 //
1502 // Found with:
1503 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1504 case Op_Phi:
1505 return false;
1506
1507 // MemBarNode::Ideal does "Eliminate volatile MemBars for scalar replaced objects".
1508 // For examle "The allocated object does not escape".
1509 //
1510 // It seems the difference to earlier calls to MemBarNode::Ideal, is that there
1511 // alloc->as_Allocate()->does_not_escape_thread() returned false, but in verification
1512 // it returned true. Why does the MemBarStoreStore not get added to the IGVN
1513 // worklist when this change happens?
1514 //
1515 // Found with:
1516 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1517 case Op_MemBarStoreStore:
1518 return false;
1519
1520 // ConvI2LNode::Ideal converts
1521 // 648 AddI === _ 583 645 [[ 661 ]]
1522 // 661 ConvI2L === _ 648 [[ 664 ]] #long:0..maxint-1:www
1523 // into
1524 // 772 ConvI2L === _ 645 [[ 773 ]] #long:-120..maxint-61:www
1525 // 771 ConvI2L === _ 583 [[ 773 ]] #long:60..120:www
1526 // 773 AddL === _ 771 772 [[ ]]
1527 //
1528 // We have to investigate why this does not happen during IGVN in this case.
1529 // There could also be other issues - I did not investigate further yet.
1530 //
1531 // Found with:
1532 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1533 case Op_ConvI2L:
1534 return false;
1535
1536 // AddNode::IdealIL can do this transform (and similar other ones):
1537 // Convert "a*b+a*c into a*(b+c)
1538 // The example had AddI(MulI(a, b), MulI(a, c)). Why did this not happen
1539 // during IGVN? There was a mutation for one of the MulI, and only
1540 // after that the pattern was as needed for the optimization. The MulI
1541 // was added to the IGVN worklist, but not the AddI. This probably
1542 // can be fixed by adding the correct pattern in add_users_of_use_to_worklist.
1543 //
1544 // Found with:
1545 // test/hotspot/jtreg/compiler/loopopts/superword/ReductionPerf.java
1546 // -XX:VerifyIterativeGVN=1110
1547 case Op_AddI:
1548 return false;
1549
1550 // ArrayCopyNode::Ideal
1551 // calls ArrayCopyNode::prepare_array_copy
1552 // calls Compile::conv_I2X_index -> is called with sizetype = intcon(0), I think that
1553 // is not expected, and we create a range int:0..-1
1554 // calls Compile::constrained_convI2L -> creates ConvI2L(intcon(1), int:0..-1)
1555 // note: the type is already empty!
1556 // calls PhaseIterGVN::transform
1557 // calls PhaseIterGVN::transform_old
1558 // calls PhaseIterGVN::subsume_node -> subsume ConvI2L with TOP
1559 // calls Unique_Node_List::push -> pushes TOP to worklist
1560 //
1561 // Once we get back to ArrayCopyNode::prepare_array_copy, we get back TOP, and
1562 // return false. This means we eventually return nullptr from ArrayCopyNode::Ideal.
1563 //
1564 // Question: is it ok to push anything to the worklist during ::Ideal, if we will
1565 // return nullptr, indicating nothing happened?
1566 // Is it smart to do transform in Compile::constrained_convI2L, and then
1567 // check for TOP in calls ArrayCopyNode::prepare_array_copy?
1568 // Should we just allow TOP to land on the worklist, as an exception?
1569 //
1570 // Found with:
1571 // compiler/arraycopy/TestArrayCopyAsLoadsStores.java
1572 // -XX:VerifyIterativeGVN=1110
1573 case Op_ArrayCopy:
1574 return false;
1575
1576 // CastLLNode::Ideal
1577 // calls ConstraintCastNode::optimize_integer_cast -> pushes CastLL through SubL
1578 //
1579 // Could be a notification issue, where updates inputs of CastLL do not notify
1580 // down through SubL to CastLL.
1581 //
1582 // Found With:
1583 // compiler/c2/TestMergeStoresMemorySegment.java#byte-array
1584 // -XX:VerifyIterativeGVN=1110
1585 case Op_CastLL:
1586 return false;
1587
1588 // Similar case happens to CastII
1589 //
1590 // Found With:
1591 // compiler/c2/TestScalarReplacementMaxLiveNodes.java
1592 // -XX:VerifyIterativeGVN=1110
1593 case Op_CastII:
1594 return false;
1595
1596 // MaxLNode::Ideal
1597 // calls AddNode::Ideal
1598 // calls commute -> decides to swap edges
1599 //
1600 // Another notification issue, because we check inputs of inputs?
1601 // MaxL -> Phi -> Loop
1602 // MaxL -> Phi -> MaxL
1603 //
1604 // Found with:
1605 // compiler/c2/irTests/TestIfMinMax.java
1606 // -XX:VerifyIterativeGVN=1110
1607 case Op_MaxL:
1608 case Op_MinL:
1609 return false;
1610
1611 // OrINode::Ideal
1612 // calls AddNode::Ideal
1613 // calls commute -> left is Load, right not -> commute.
1614 //
1615 // Not sure why notification does not work here, seems like
1616 // the depth is only 1, so it should work. Needs investigation.
1617 //
1618 // Found with:
1619 // compiler/codegen/TestCharVect2.java#id0
1620 // -XX:VerifyIterativeGVN=1110
1621 case Op_OrI:
1622 case Op_OrL:
1623 return false;
1624
1625 // Bool -> constant folded to 1.
1626 // Issue with notification?
1627 //
1628 // Found with:
1629 // compiler/c2/irTests/TestVectorizationMismatchedAccess.java
1630 // -XX:VerifyIterativeGVN=1110
1631 case Op_Bool:
1632 return false;
1633
1634 // LShiftLNode::Ideal
1635 // Looks at pattern: "(x + x) << c0", converts it to "x << (c0 + 1)"
1636 // Probably a notification issue.
1637 //
1638 // Found with:
1639 // compiler/conversions/TestMoveConvI2LOrCastIIThruAddIs.java
1640 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1641 case Op_LShiftL:
1642 return false;
1643
1644 // LShiftINode::Ideal
1645 // pattern: ((x + con1) << con2) -> x << con2 + con1 << con2
1646 // Could be issue with notification of inputs of inputs
1647 //
1648 // Side-note: should cases like these not be shared between
1649 // LShiftI and LShiftL?
1650 //
1651 // Found with:
1652 // compiler/escapeAnalysis/Test6689060.java
1653 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110 -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1654 case Op_LShiftI:
1655 return false;
1656
1657 // AddPNode::Ideal seems to do set_req without removing lock first.
1658 // Found with various vector tests tier1-tier3.
1659 case Op_AddP:
1660 return false;
1661
1662 // StrIndexOfNode::Ideal
1663 // Found in tier1-3.
1664 case Op_StrIndexOf:
1665 case Op_StrIndexOfChar:
1666 return false;
1667
1668 // StrEqualsNode::Identity
1669 //
1670 // Found (linux x64 only?) with:
1671 // serviceability/sa/ClhsdbThreadContext.java
1672 // -XX:+UnlockExperimentalVMOptions -XX:LockingMode=1 -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1673 case Op_StrEquals:
1674 return false;
1675
1676 // AryEqNode::Ideal
1677 // Not investigated. Reshapes itself and adds lots of nodes to the worklist.
1678 //
1679 // Found with:
1680 // vmTestbase/vm/mlvm/meth/stress/compiler/i2c_c2i/Test.java
1681 // -XX:+UnlockDiagnosticVMOptions -XX:-TieredCompilation -XX:+StressUnstableIfTraps -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1682 case Op_AryEq:
1683 return false;
1684
1685 // MergeMemNode::Ideal
1686 // Found in tier1-3. Did not investigate further yet.
1687 case Op_MergeMem:
1688 return false;
1689
1690 // URShiftINode::Ideal
1691 // Found in tier1-3. Did not investigate further yet.
1692 case Op_URShiftI:
1693 return false;
1694
1695 // CMoveINode::Ideal
1696 // Found in tier1-3. Did not investigate further yet.
1697 case Op_CMoveI:
1698 return false;
1699
1700 // CmpPNode::Ideal calls isa_const_java_mirror
1701 // and generates new constant nodes, even if no progress is made.
1702 // We can probably rewrite this so that only types are generated.
1703 // It seems that object types are not hashed, we could investigate
1704 // if that is an option as well.
1705 //
1706 // Found with:
1707 // java -XX:VerifyIterativeGVN=1110 -Xcomp --version
1708 case Op_CmpP:
1709 return false;
1710
1711 // MinINode::Ideal
1712 // Did not investigate, but there are some patterns that might
1713 // need more notification.
1714 case Op_MinI:
1715 case Op_MaxI: // preemptively removed it as well.
1716 return false;
1717 }
1718
1719 if (n->is_Load()) {
1720 // LoadNode::Ideal uses tries to find an earlier memory state, and
1721 // checks can_see_stored_value for it.
1722 //
1723 // Investigate why this was not already done during IGVN.
1724 // A similar issue happens with Identity.
1725 //
1726 // There seem to be other cases where loads go up some steps, like
1727 // LoadNode::Ideal going up 10x steps to find dominating load.
1728 //
1729 // Found with:
1730 // test/hotspot/jtreg/compiler/arraycopy/TestCloneAccess.java
1731 // -XX:VerifyIterativeGVN=1110
1732 return false;
1733 }
1734
1735 if (n->is_Store()) {
1736 // StoreNode::Ideal can do this:
1737 // // Capture an unaliased, unconditional, simple store into an initializer.
1738 // // Or, if it is independent of the allocation, hoist it above the allocation.
1739 // That replaces the Store with a MergeMem.
1740 //
1741 // We have to investigate why this does not happen during IGVN in this case.
1742 // There could also be other issues - I did not investigate further yet.
1743 //
1744 // Found with:
1745 // java -XX:VerifyIterativeGVN=0100 -Xcomp --version
1746 return false;
1747 }
1748
1749 if (n->is_Vector()) {
1750 // VectorNode::Ideal swaps edges, but only for ops
1751 // that are deemed commutable. But swap_edges
1752 // requires the hash to be invariant when the edges
1753 // are swapped, which is not implemented for these
1754 // vector nodes. This seems not to create any trouble
1755 // usually, but we can also get graphs where in the
1756 // end the nodes are not all commuted, so there is
1757 // definitively an issue here.
1758 //
1759 // Probably we have two options: kill the hash, or
1760 // properly make the hash commutation friendly.
1761 //
1762 // Found with:
1763 // compiler/vectorapi/TestMaskedMacroLogicVector.java
1764 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110 -XX:+UseParallelGC -XX:+UseNUMA
1765 return false;
1766 }
1767
1768 if (n->is_Region()) {
1769 // LoopNode::Ideal calls RegionNode::Ideal.
1770 // CountedLoopNode::Ideal calls RegionNode::Ideal too.
1771 // But I got an issue because RegionNode::optimize_trichotomy
1772 // then modifies another node, and pushes nodes to the worklist
1773 // Not sure if this is ok, modifying another node like that.
1774 // Maybe it is, then we need to look into what to do with
1775 // the nodes that are now on the worklist, maybe just clear
1776 // them out again. But maybe modifying other nodes like that
1777 // is also bad design. In the end, we return nullptr for
1778 // the current CountedLoop. But the extra nodes on the worklist
1779 // trip the asserts later on.
1780 //
1781 // Found with:
1782 // compiler/eliminateAutobox/TestShortBoxing.java
1783 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1784 return false;
1785 }
1786
1787 if (n->is_CallJava()) {
1788 // CallStaticJavaNode::Ideal
1789 // Led to a crash:
1790 // assert((is_CallStaticJava() && cg->is_mh_late_inline()) || (is_CallDynamicJava() && cg->is_virtual_late_inline())) failed: mismatch
1791 //
1792 // Did not investigate yet, could be a bug.
1793 // Or maybe it does not expect to be called during verification.
1794 //
1795 // Found with:
1796 // test/jdk/jdk/incubator/vector/VectorRuns.java
1797 // -XX:VerifyIterativeGVN=1110
1798
1799 // CallDynamicJavaNode::Ideal, and I think also for CallStaticJavaNode::Ideal
1800 // and possibly their subclasses.
1801 // During late inlining it can call CallJavaNode::register_for_late_inline
1802 // That means we do more rounds of late inlining, but might fail.
1803 // Then we do IGVN again, and register the node again for late inlining.
1804 // This creates an endless cycle. Everytime we try late inlining, we
1805 // are also creating more nodes, especially SafePoint and MergeMem.
1806 // These nodes are immediately rejected when the inlining fails in the
1807 // do_late_inline_check, but they still grow the memory, until we hit
1808 // the MemLimit and crash.
1809 // The assumption here seems that CallDynamicJavaNode::Ideal does not get
1810 // called repeatedly, and eventually we terminate. I fear this is not
1811 // a great assumption to make. We should investigate more.
1812 //
1813 // Found with:
1814 // compiler/loopopts/superword/TestDependencyOffsets.java#vanilla-U
1815 // -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1816 return false;
1817 }
1818
1819 // The number of nodes shoud not increase.
1820 uint old_unique = C->unique();
1821
1822 Node* i = n->Ideal(this, can_reshape);
1823 // If there was no new Idealization, we are probably happy.
1824 if (i == nullptr) {
1825 if (old_unique < C->unique()) {
1826 stringStream ss; // Print as a block without tty lock.
1827 ss.cr();
1828 ss.print_cr("Ideal optimization did not make progress but created new unused nodes.");
1829 ss.print_cr(" old_unique = %d, unique = %d", old_unique, C->unique());
1830 n->dump_bfs(1, nullptr, "", &ss);
1831 tty->print_cr("%s", ss.as_string());
1832 return true;
1833 }
1834
1835 verify_empty_worklist(n);
1836
1837 // Everything is good.
1838 return false;
1839 }
1840
1841 // We just saw a new Idealization which was not done during IGVN.
1842 stringStream ss; // Print as a block without tty lock.
1843 ss.cr();
1844 ss.print_cr("Missed Ideal optimization (can_reshape=%s):", can_reshape ? "true": "false");
1845 if (i == n) {
1846 ss.print_cr("The node was reshaped by Ideal.");
1847 } else {
1848 ss.print_cr("The node was replaced by Ideal.");
1849 ss.print_cr("Old node:");
1850 n->dump_bfs(1, nullptr, "", &ss);
1851 }
1852 ss.print_cr("The result after Ideal:");
1853 i->dump_bfs(1, nullptr, "", &ss);
1854 tty->print_cr("%s", ss.as_string());
1855 return true;
1856 }
1857
1858 // Check that all Identity optimizations that could be done were done.
1859 // Returns true if it found missed optimization opportunities and
1860 // false otherwise (no missed optimization, or skipped verification).
1861 bool PhaseIterGVN::verify_Identity_for(Node* n) {
1862 // First, we check a list of exceptions, where we skip verification,
1863 // because there are known cases where Ideal can optimize after IGVN.
1864 // Some may be expected and cannot be fixed, and others should be fixed.
1865 switch (n->Opcode()) {
1866 // SafePointNode::Identity can remove SafePoints, but wants to wait until
1867 // after loopopts:
1868 // // Transforming long counted loops requires a safepoint node. Do not
1869 // // eliminate a safepoint until loop opts are over.
1870 // if (in(0)->is_Proj() && !phase->C->major_progress()) {
1871 //
1872 // I think the check for major_progress does delay it until after loopopts
1873 // but it does not ensure that the node is on the IGVN worklist after
1874 // loopopts. I think we should try to instead check for
1875 // phase->C->post_loop_opts_phase() and call record_for_post_loop_opts_igvn.
1876 //
1877 // Found with:
1878 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1879 case Op_SafePoint:
1880 return false;
1881
1882 // MergeMemNode::Identity replaces the MergeMem with its base_memory if it
1883 // does not record any other memory splits.
1884 //
1885 // I did not deeply investigate, but it looks like MergeMemNode::Identity
1886 // never got called during IGVN for this node, investigate why.
1887 //
1888 // Found with:
1889 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1890 case Op_MergeMem:
1891 return false;
1892
1893 // ConstraintCastNode::Identity finds casts that are the same, except that
1894 // the control is "higher up", i.e. dominates. The call goes via
1895 // ConstraintCastNode::dominating_cast to PhaseGVN::is_dominator_helper,
1896 // which traverses up to 100 idom steps. If anything gets optimized somewhere
1897 // away from the cast, but within 100 idom steps, the cast may not be
1898 // put on the IGVN worklist any more.
1899 //
1900 // Found with:
1901 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1902 case Op_CastPP:
1903 case Op_CastII:
1904 case Op_CastLL:
1905 return false;
1906
1907 // Same issue for CheckCastPP, uses ConstraintCastNode::Identity and
1908 // checks dominator, which may be changed, but too far up for notification
1909 // to work.
1910 //
1911 // Found with:
1912 // compiler/c2/irTests/TestSkeletonPredicates.java
1913 // -XX:VerifyIterativeGVN=1110
1914 case Op_CheckCastPP:
1915 return false;
1916
1917 // In SubNode::Identity, we do:
1918 // Convert "(X+Y) - Y" into X and "(X+Y) - X" into Y
1919 // In the example, the AddI had an input replaced, the AddI is
1920 // added to the IGVN worklist, but the SubI is one link further
1921 // down and is not added. I checked add_users_of_use_to_worklist
1922 // where I would expect the SubI would be added, and I cannot
1923 // find the pattern, only this one:
1924 // If changed AddI/SubI inputs, check CmpU for range check optimization.
1925 //
1926 // Fix this "notification" issue and check if there are any other
1927 // issues.
1928 //
1929 // Found with:
1930 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
1931 case Op_SubI:
1932 case Op_SubL:
1933 return false;
1934
1935 // PhiNode::Identity checks for patterns like:
1936 // r = (x != con) ? x : con;
1937 // that can be constant folded to "x".
1938 //
1939 // Call goes through PhiNode::is_cmove_id and CMoveNode::is_cmove_id.
1940 // I suspect there was some earlier change to one of the inputs, but
1941 // not all relevant outputs were put on the IGVN worklist.
1942 //
1943 // Found with:
1944 // test/hotspot/jtreg/gc/stress/gcbasher/TestGCBasherWithG1.java
1945 // -XX:VerifyIterativeGVN=1110
1946 case Op_Phi:
1947 return false;
1948
1949 // ConvI2LNode::Identity does
1950 // convert I2L(L2I(x)) => x
1951 //
1952 // Investigate why this did not already happen during IGVN.
1953 //
1954 // Found with:
1955 // compiler/loopopts/superword/TestDependencyOffsets.java#vanilla-A
1956 // -XX:VerifyIterativeGVN=1110
1957 case Op_ConvI2L:
1958 return false;
1959
1960 // MaxNode::find_identity_operation
1961 // Finds patterns like Max(A, Max(A, B)) -> Max(A, B)
1962 // This can be a 2-hop search, so maybe notification is not
1963 // good enough.
1964 //
1965 // Found with:
1966 // compiler/codegen/TestBooleanVect.java
1967 // -XX:VerifyIterativeGVN=1110
1968 case Op_MaxL:
1969 case Op_MinL:
1970 case Op_MaxI:
1971 case Op_MinI:
1972 case Op_MaxF:
1973 case Op_MinF:
1974 case Op_MaxHF:
1975 case Op_MinHF:
1976 case Op_MaxD:
1977 case Op_MinD:
1978 return false;
1979
1980
1981 // AddINode::Identity
1982 // Converts (x-y)+y to x
1983 // Could be issue with notification
1984 //
1985 // Turns out AddL does the same.
1986 //
1987 // Found with:
1988 // compiler/c2/Test6792161.java
1989 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
1990 case Op_AddI:
1991 case Op_AddL:
1992 return false;
1993
1994 // AbsINode::Identity
1995 // Not investigated yet.
1996 case Op_AbsI:
1997 return false;
1998 }
1999
2000 if (n->is_Load()) {
2001 // LoadNode::Identity tries to look for an earlier store value via
2002 // can_see_stored_value. I found an example where this led to
2003 // an Allocation, where we could assume the value was still zero.
2004 // So the LoadN can be replaced with a zerocon.
2005 //
2006 // Investigate why this was not already done during IGVN.
2007 // A similar issue happens with Ideal.
2008 //
2009 // Found with:
2010 // java -XX:VerifyIterativeGVN=1000 -Xcomp --version
2011 return false;
2012 }
2013
2014 if (n->is_Store()) {
2015 // StoreNode::Identity
2016 // Not investigated, but found missing optimization for StoreI.
2017 // Looks like a StoreI is replaced with an InitializeNode.
2018 //
2019 // Found with:
2020 // applications/ctw/modules/java_base_2.java
2021 // -ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation -Djava.awt.headless=true -XX:+IgnoreUnrecognizedVMOptions -XX:VerifyIterativeGVN=1110
2022 return false;
2023 }
2024
2025 if (n->is_Vector()) {
2026 // Found with tier1-3. Not investigated yet.
2027 // The observed issue was with AndVNode::Identity
2028 return false;
2029 }
2030
2031 Node* i = n->Identity(this);
2032 // If we cannot find any other Identity, we are happy.
2033 if (i == n) {
2034 verify_empty_worklist(n);
2035 return false;
2036 }
2037
2038 // The verification just found a new Identity that was not found during IGVN.
2039 stringStream ss; // Print as a block without tty lock.
2040 ss.cr();
2041 ss.print_cr("Missed Identity optimization:");
2042 ss.print_cr("Old node:");
2043 n->dump_bfs(1, nullptr, "", &ss);
2044 ss.print_cr("New node:");
2045 i->dump_bfs(1, nullptr, "", &ss);
2046 tty->print_cr("%s", ss.as_string());
2047 return true;
2048 }
2049 #endif
2050
2051 /**
2052 * Register a new node with the optimizer. Update the types array, the def-use
2053 * info. Put on worklist.
2054 */
2055 Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) {
2056 set_type_bottom(n);
2057 _worklist.push(n);
2058 if (orig != nullptr) C->copy_node_notes_to(n, orig);
2059 return n;
2060 }
2061
2062 //------------------------------transform--------------------------------------
2063 // Non-recursive: idealize Node 'n' with respect to its inputs and its value
2064 Node *PhaseIterGVN::transform( Node *n ) {
2065 // If brand new node, make space in type array, and give it a type.
2066 ensure_type_or_null(n);
2067 if (type_or_null(n) == nullptr) {
2068 set_type_bottom(n);
2069 }
2070
2071 if (_delay_transform) {
2072 // Add the node to the worklist but don't optimize for now
2073 _worklist.push(n);
2074 return n;
2075 }
2076
2077 return transform_old(n);
2078 }
2079
2080 Node *PhaseIterGVN::transform_old(Node* n) {
2081 NOT_PRODUCT(set_transforms());
2082 // Remove 'n' from hash table in case it gets modified
2083 _table.hash_delete(n);
2084 #ifdef ASSERT
2085 if (is_verify_def_use()) {
2086 assert(!_table.find_index(n->_idx), "found duplicate entry in table");
2087 }
2088 #endif
2089
2090 // Allow Bool -> Cmp idealisation in late inlining intrinsics that return a bool
2091 if (n->is_Cmp()) {
2092 add_users_to_worklist(n);
2093 }
2094
2095 // Apply the Ideal call in a loop until it no longer applies
2096 Node* k = n;
2097 DEBUG_ONLY(dead_loop_check(k);)
2098 DEBUG_ONLY(bool is_new = (k->outcnt() == 0);)
2099 C->remove_modified_node(k);
2100 Node* i = apply_ideal(k, /*can_reshape=*/true);
2101 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes");
2102 #ifndef PRODUCT
2103 verify_step(k);
2104 #endif
2105
2106 DEBUG_ONLY(uint loop_count = 1;)
2107 while (i != nullptr) {
2108 #ifdef ASSERT
2109 if (loop_count >= K + C->live_nodes()) {
2110 dump_infinite_loop_info(i, "PhaseIterGVN::transform_old");
2111 }
2112 #endif
2113 assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes");
2114 // Made a change; put users of original Node on worklist
2115 add_users_to_worklist(k);
2116 // Replacing root of transform tree?
2117 if (k != i) {
2118 // Make users of old Node now use new.
2119 subsume_node(k, i);
2120 k = i;
2121 }
2122 DEBUG_ONLY(dead_loop_check(k);)
2123 // Try idealizing again
2124 DEBUG_ONLY(is_new = (k->outcnt() == 0);)
2125 C->remove_modified_node(k);
2126 i = apply_ideal(k, /*can_reshape=*/true);
2127 assert(i != k || is_new || (i->outcnt() > 0), "don't return dead nodes");
2128 #ifndef PRODUCT
2129 verify_step(k);
2130 #endif
2131 DEBUG_ONLY(loop_count++;)
2132 }
2133
2134 // If brand new node, make space in type array.
2135 ensure_type_or_null(k);
2136
2137 // See what kind of values 'k' takes on at runtime
2138 const Type* t = k->Value(this);
2139 assert(t != nullptr, "value sanity");
2140
2141 // Since I just called 'Value' to compute the set of run-time values
2142 // for this Node, and 'Value' is non-local (and therefore expensive) I'll
2143 // cache Value. Later requests for the local phase->type of this Node can
2144 // use the cached Value instead of suffering with 'bottom_type'.
2145 if (type_or_null(k) != t) {
2146 #ifndef PRODUCT
2147 inc_new_values();
2148 set_progress();
2149 #endif
2150 set_type(k, t);
2151 // If k is a TypeNode, capture any more-precise type permanently into Node
2152 k->raise_bottom_type(t);
2153 // Move users of node to worklist
2154 add_users_to_worklist(k);
2155 }
2156 // If 'k' computes a constant, replace it with a constant
2157 if (t->singleton() && !k->is_Con()) {
2158 NOT_PRODUCT(set_progress();)
2159 Node* con = makecon(t); // Make a constant
2160 add_users_to_worklist(k);
2161 subsume_node(k, con); // Everybody using k now uses con
2162 return con;
2163 }
2164
2165 // Now check for Identities
2166 i = k->Identity(this); // Look for a nearby replacement
2167 if (i != k) { // Found? Return replacement!
2168 NOT_PRODUCT(set_progress();)
2169 add_users_to_worklist(k);
2170 subsume_node(k, i); // Everybody using k now uses i
2171 return i;
2172 }
2173
2174 // Global Value Numbering
2175 i = hash_find_insert(k); // Check for pre-existing node
2176 if (i && (i != k)) {
2177 // Return the pre-existing node if it isn't dead
2178 NOT_PRODUCT(set_progress();)
2179 add_users_to_worklist(k);
2180 subsume_node(k, i); // Everybody using k now uses i
2181 return i;
2182 }
2183
2184 // Return Idealized original
2185 return k;
2186 }
2187
2188 //---------------------------------saturate------------------------------------
2189 const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type,
2190 const Type* limit_type) const {
2191 return new_type->narrow(old_type);
2192 }
2193
2194 //------------------------------remove_globally_dead_node----------------------
2195 // Kill a globally dead Node. All uses are also globally dead and are
2196 // aggressively trimmed.
2197 void PhaseIterGVN::remove_globally_dead_node( Node *dead ) {
2198 enum DeleteProgress {
2199 PROCESS_INPUTS,
2200 PROCESS_OUTPUTS
2201 };
2202 ResourceMark rm;
2203 Node_Stack stack(32);
2204 stack.push(dead, PROCESS_INPUTS);
2205
2206 while (stack.is_nonempty()) {
2207 dead = stack.node();
2208 if (dead->Opcode() == Op_SafePoint) {
2209 dead->as_SafePoint()->disconnect_from_root(this);
2210 }
2211 uint progress_state = stack.index();
2212 assert(dead != C->root(), "killing root, eh?");
2213 assert(!dead->is_top(), "add check for top when pushing");
2214 NOT_PRODUCT( set_progress(); )
2215 if (progress_state == PROCESS_INPUTS) {
2216 // After following inputs, continue to outputs
2217 stack.set_index(PROCESS_OUTPUTS);
2218 if (!dead->is_Con()) { // Don't kill cons but uses
2219 bool recurse = false;
2220 // Remove from hash table
2221 _table.hash_delete( dead );
2222 // Smash all inputs to 'dead', isolating him completely
2223 for (uint i = 0; i < dead->req(); i++) {
2224 Node *in = dead->in(i);
2225 if (in != nullptr && in != C->top()) { // Points to something?
2226 int nrep = dead->replace_edge(in, nullptr, this); // Kill edges
2227 assert((nrep > 0), "sanity");
2228 if (in->outcnt() == 0) { // Made input go dead?
2229 stack.push(in, PROCESS_INPUTS); // Recursively remove
2230 recurse = true;
2231 } else if (in->outcnt() == 1 &&
2232 in->has_special_unique_user()) {
2233 _worklist.push(in->unique_out());
2234 } else if (in->outcnt() <= 2 && dead->is_Phi()) {
2235 if (in->Opcode() == Op_Region) {
2236 _worklist.push(in);
2237 } else if (in->is_Store()) {
2238 DUIterator_Fast imax, i = in->fast_outs(imax);
2239 _worklist.push(in->fast_out(i));
2240 i++;
2241 if (in->outcnt() == 2) {
2242 _worklist.push(in->fast_out(i));
2243 i++;
2244 }
2245 assert(!(i < imax), "sanity");
2246 }
2247 } else if (dead->is_data_proj_of_pure_function(in)) {
2248 _worklist.push(in);
2249 } else {
2250 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(this, in);
2251 }
2252 if (ReduceFieldZeroing && dead->is_Load() && i == MemNode::Memory &&
2253 in->is_Proj() && in->in(0) != nullptr && in->in(0)->is_Initialize()) {
2254 // A Load that directly follows an InitializeNode is
2255 // going away. The Stores that follow are candidates
2256 // again to be captured by the InitializeNode.
2257 for (DUIterator_Fast jmax, j = in->fast_outs(jmax); j < jmax; j++) {
2258 Node *n = in->fast_out(j);
2259 if (n->is_Store()) {
2260 _worklist.push(n);
2261 }
2262 }
2263 }
2264 } // if (in != nullptr && in != C->top())
2265 } // for (uint i = 0; i < dead->req(); i++)
2266 if (recurse) {
2267 continue;
2268 }
2269 } // if (!dead->is_Con())
2270 } // if (progress_state == PROCESS_INPUTS)
2271
2272 // Aggressively kill globally dead uses
2273 // (Rather than pushing all the outs at once, we push one at a time,
2274 // plus the parent to resume later, because of the indefinite number
2275 // of edge deletions per loop trip.)
2276 if (dead->outcnt() > 0) {
2277 // Recursively remove output edges
2278 stack.push(dead->raw_out(0), PROCESS_INPUTS);
2279 } else {
2280 // Finished disconnecting all input and output edges.
2281 stack.pop();
2282 // Remove dead node from iterative worklist
2283 _worklist.remove(dead);
2284 C->remove_useless_node(dead);
2285 }
2286 } // while (stack.is_nonempty())
2287 }
2288
2289 //------------------------------subsume_node-----------------------------------
2290 // Remove users from node 'old' and add them to node 'nn'.
2291 void PhaseIterGVN::subsume_node( Node *old, Node *nn ) {
2292 if (old->Opcode() == Op_SafePoint) {
2293 old->as_SafePoint()->disconnect_from_root(this);
2294 }
2295 assert( old != hash_find(old), "should already been removed" );
2296 assert( old != C->top(), "cannot subsume top node");
2297 // Copy debug or profile information to the new version:
2298 C->copy_node_notes_to(nn, old);
2299 // Move users of node 'old' to node 'nn'
2300 for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) {
2301 Node* use = old->last_out(i); // for each use...
2302 // use might need re-hashing (but it won't if it's a new node)
2303 rehash_node_delayed(use);
2304 // Update use-def info as well
2305 // We remove all occurrences of old within use->in,
2306 // so as to avoid rehashing any node more than once.
2307 // The hash table probe swamps any outer loop overhead.
2308 uint num_edges = 0;
2309 for (uint jmax = use->len(), j = 0; j < jmax; j++) {
2310 if (use->in(j) == old) {
2311 use->set_req(j, nn);
2312 ++num_edges;
2313 }
2314 }
2315 i -= num_edges; // we deleted 1 or more copies of this edge
2316 }
2317
2318 // Search for instance field data PhiNodes in the same region pointing to the old
2319 // memory PhiNode and update their instance memory ids to point to the new node.
2320 if (old->is_Phi() && old->as_Phi()->type()->has_memory() && old->in(0) != nullptr) {
2321 Node* region = old->in(0);
2322 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2323 PhiNode* phi = region->fast_out(i)->isa_Phi();
2324 if (phi != nullptr && phi->inst_mem_id() == (int)old->_idx) {
2325 phi->set_inst_mem_id((int)nn->_idx);
2326 }
2327 }
2328 }
2329
2330 // Smash all inputs to 'old', isolating him completely
2331 Node *temp = new Node(1);
2332 temp->init_req(0,nn); // Add a use to nn to prevent him from dying
2333 remove_dead_node( old );
2334 temp->del_req(0); // Yank bogus edge
2335 if (nn != nullptr && nn->outcnt() == 0) {
2336 _worklist.push(nn);
2337 }
2338 #ifndef PRODUCT
2339 if (is_verify_def_use()) {
2340 for ( int i = 0; i < _verify_window_size; i++ ) {
2341 if ( _verify_window[i] == old )
2342 _verify_window[i] = nn;
2343 }
2344 }
2345 #endif
2346 temp->destruct(this); // reuse the _idx of this little guy
2347 }
2348
2349 void PhaseIterGVN::replace_in_uses(Node* n, Node* m) {
2350 assert(n != nullptr, "sanity");
2351 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2352 Node* u = n->fast_out(i);
2353 if (u != n) {
2354 rehash_node_delayed(u);
2355 int nb = u->replace_edge(n, m);
2356 --i, imax -= nb;
2357 }
2358 }
2359 assert(n->outcnt() == 0, "all uses must be deleted");
2360 }
2361
2362 //------------------------------add_users_to_worklist--------------------------
2363 void PhaseIterGVN::add_users_to_worklist0(Node* n, Unique_Node_List& worklist) {
2364 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2365 worklist.push(n->fast_out(i)); // Push on worklist
2366 }
2367 }
2368
2369 // Return counted loop Phi if as a counted loop exit condition, cmp
2370 // compares the induction variable with n
2371 static PhiNode* countedloop_phi_from_cmp(CmpNode* cmp, Node* n) {
2372 for (DUIterator_Fast imax, i = cmp->fast_outs(imax); i < imax; i++) {
2373 Node* bol = cmp->fast_out(i);
2374 for (DUIterator_Fast i2max, i2 = bol->fast_outs(i2max); i2 < i2max; i2++) {
2375 Node* iff = bol->fast_out(i2);
2376 if (iff->is_BaseCountedLoopEnd()) {
2377 BaseCountedLoopEndNode* cle = iff->as_BaseCountedLoopEnd();
2378 if (cle->limit() == n) {
2379 PhiNode* phi = cle->phi();
2380 if (phi != nullptr) {
2381 return phi;
2382 }
2383 }
2384 }
2385 }
2386 }
2387 return nullptr;
2388 }
2389
2390 void PhaseIterGVN::add_users_to_worklist(Node *n) {
2391 add_users_to_worklist0(n, _worklist);
2392
2393 Unique_Node_List& worklist = _worklist;
2394 // Move users of node to worklist
2395 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2396 Node* use = n->fast_out(i); // Get use
2397 add_users_of_use_to_worklist(n, use, worklist);
2398 }
2399 }
2400
2401 void PhaseIterGVN::add_users_of_use_to_worklist(Node* n, Node* use, Unique_Node_List& worklist) {
2402 if(use->is_Multi() || // Multi-definer? Push projs on worklist
2403 use->is_Store() ) // Enable store/load same address
2404 add_users_to_worklist0(use, worklist);
2405
2406 // If we changed the receiver type to a call, we need to revisit
2407 // the Catch following the call. It's looking for a non-null
2408 // receiver to know when to enable the regular fall-through path
2409 // in addition to the NullPtrException path.
2410 if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) {
2411 Node* p = use->as_CallDynamicJava()->proj_out_or_null(TypeFunc::Control);
2412 if (p != nullptr) {
2413 add_users_to_worklist0(p, worklist);
2414 }
2415 }
2416
2417 // AndLNode::Ideal folds GraphKit::mark_word_test patterns. Give it a chance to run.
2418 if (n->is_Load() && use->is_Phi()) {
2419 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
2420 Node* u = use->fast_out(i);
2421 if (u->Opcode() == Op_AndL) {
2422 worklist.push(u);
2423 }
2424 }
2425 }
2426
2427 uint use_op = use->Opcode();
2428 if(use->is_Cmp()) { // Enable CMP/BOOL optimization
2429 add_users_to_worklist0(use, worklist); // Put Bool on worklist
2430 if (use->outcnt() > 0) {
2431 Node* bol = use->raw_out(0);
2432 if (bol->outcnt() > 0) {
2433 Node* iff = bol->raw_out(0);
2434 if (iff->outcnt() == 2) {
2435 // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the
2436 // phi merging either 0 or 1 onto the worklist
2437 Node* ifproj0 = iff->raw_out(0);
2438 Node* ifproj1 = iff->raw_out(1);
2439 if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) {
2440 Node* region0 = ifproj0->raw_out(0);
2441 Node* region1 = ifproj1->raw_out(0);
2442 if( region0 == region1 )
2443 add_users_to_worklist0(region0, worklist);
2444 }
2445 }
2446 }
2447 }
2448 if (use_op == Op_CmpI || use_op == Op_CmpL) {
2449 Node* phi = countedloop_phi_from_cmp(use->as_Cmp(), n);
2450 if (phi != nullptr) {
2451 // Input to the cmp of a loop exit check has changed, thus
2452 // the loop limit may have changed, which can then change the
2453 // range values of the trip-count Phi.
2454 worklist.push(phi);
2455 }
2456 }
2457 if (use_op == Op_CmpI) {
2458 Node* cmp = use;
2459 Node* in1 = cmp->in(1);
2460 Node* in2 = cmp->in(2);
2461 // Notify CmpI / If pattern from CastIINode::Value (left pattern).
2462 // Must also notify if in1 is modified and possibly turns into X (right pattern).
2463 //
2464 // in1 in2 in1 in2
2465 // | | | |
2466 // +--- | --+ | |
2467 // | | | | |
2468 // CmpINode | CmpINode
2469 // | | |
2470 // BoolNode | BoolNode
2471 // | | OR |
2472 // IfNode | IfNode
2473 // | | |
2474 // IfProj | IfProj X
2475 // | | | |
2476 // CastIINode CastIINode
2477 //
2478 if (in1 != in2) { // if they are equal, the CmpI can fold them away
2479 if (in1 == n) {
2480 // in1 modified -> could turn into X -> do traversal based on right pattern.
2481 for (DUIterator_Fast i2max, i2 = cmp->fast_outs(i2max); i2 < i2max; i2++) {
2482 Node* bol = cmp->fast_out(i2); // For each Bool
2483 if (bol->is_Bool()) {
2484 for (DUIterator_Fast i3max, i3 = bol->fast_outs(i3max); i3 < i3max; i3++) {
2485 Node* iff = bol->fast_out(i3); // For each If
2486 if (iff->is_If()) {
2487 for (DUIterator_Fast i4max, i4 = iff->fast_outs(i4max); i4 < i4max; i4++) {
2488 Node* if_proj = iff->fast_out(i4); // For each IfProj
2489 assert(if_proj->is_IfProj(), "If only has IfTrue and IfFalse as outputs");
2490 for (DUIterator_Fast i5max, i5 = if_proj->fast_outs(i5max); i5 < i5max; i5++) {
2491 Node* castii = if_proj->fast_out(i5); // For each CastII
2492 if (castii->is_CastII() &&
2493 castii->as_CastII()->carry_dependency()) {
2494 worklist.push(castii);
2495 }
2496 }
2497 }
2498 }
2499 }
2500 }
2501 }
2502 } else {
2503 // Only in2 modified -> can assume X == in2 (left pattern).
2504 assert(n == in2, "only in2 modified");
2505 // Find all CastII with input in1.
2506 for (DUIterator_Fast jmax, j = in1->fast_outs(jmax); j < jmax; j++) {
2507 Node* castii = in1->fast_out(j);
2508 if (castii->is_CastII() && castii->as_CastII()->carry_dependency()) {
2509 // Find If.
2510 if (castii->in(0) != nullptr && castii->in(0)->in(0) != nullptr && castii->in(0)->in(0)->is_If()) {
2511 Node* ifnode = castii->in(0)->in(0);
2512 // Check that if connects to the cmp
2513 if (ifnode->in(1) != nullptr && ifnode->in(1)->is_Bool() && ifnode->in(1)->in(1) == cmp) {
2514 worklist.push(castii);
2515 }
2516 }
2517 }
2518 }
2519 }
2520 }
2521 }
2522 }
2523
2524 // Inline type nodes can have other inline types as users. If an input gets
2525 // updated, make sure that inline type users get a chance for optimization.
2526 if (use->is_InlineType()) {
2527 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2528 Node* u = use->fast_out(i2);
2529 if (u->is_InlineType())
2530 worklist.push(u);
2531 }
2532 }
2533 // If changed Cast input, notify down for Phi, Sub, and Xor - all do "uncast"
2534 // Patterns:
2535 // ConstraintCast+ -> Sub
2536 // ConstraintCast+ -> Phi
2537 // ConstraintCast+ -> Xor
2538 if (use->is_ConstraintCast()) {
2539 auto push_the_uses_to_worklist = [&](Node* n){
2540 if (n->is_Phi() || n->is_Sub() || n->Opcode() == Op_XorI || n->Opcode() == Op_XorL) {
2541 worklist.push(n);
2542 }
2543 };
2544 auto is_boundary = [](Node* n){ return !n->is_ConstraintCast(); };
2545 use->visit_uses(push_the_uses_to_worklist, is_boundary);
2546 }
2547 // If changed LShift inputs, check RShift users for useless sign-ext
2548 if (use_op == Op_LShiftI || use_op == Op_LShiftL) {
2549 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2550 Node* u = use->fast_out(i2);
2551 if (u->Opcode() == Op_RShiftI || u->Opcode() == Op_RShiftL)
2552 worklist.push(u);
2553 }
2554 }
2555 // If changed LShift inputs, check And users for shift and mask (And) operation
2556 if (use_op == Op_LShiftI || use_op == Op_LShiftL) {
2557 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2558 Node* u = use->fast_out(i2);
2559 if (u->Opcode() == Op_AndI || u->Opcode() == Op_AndL) {
2560 worklist.push(u);
2561 }
2562 }
2563 }
2564 // If changed AddI/SubI inputs, check CmpU for range check optimization.
2565 if (use_op == Op_AddI || use_op == Op_SubI) {
2566 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2567 Node* u = use->fast_out(i2);
2568 if (u->is_Cmp() && (u->Opcode() == Op_CmpU)) {
2569 worklist.push(u);
2570 }
2571 }
2572 }
2573 // If changed AddP inputs:
2574 // - check Stores for loop invariant, and
2575 // - if the changed input is the offset, check constant-offset AddP users for
2576 // address expression flattening.
2577 if (use_op == Op_AddP) {
2578 bool offset_changed = n == use->in(AddPNode::Offset);
2579 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2580 Node* u = use->fast_out(i2);
2581 if (u->is_Mem()) {
2582 worklist.push(u);
2583 } else if (offset_changed && u->is_AddP() && u->in(AddPNode::Offset)->is_Con()) {
2584 worklist.push(u);
2585 }
2586 }
2587 }
2588 // If changed initialization activity, check dependent Stores
2589 if (use_op == Op_Allocate || use_op == Op_AllocateArray) {
2590 InitializeNode* init = use->as_Allocate()->initialization();
2591 if (init != nullptr) {
2592 Node* imem = init->proj_out_or_null(TypeFunc::Memory);
2593 if (imem != nullptr) add_users_to_worklist0(imem, worklist);
2594 }
2595 }
2596 // If the ValidLengthTest input changes then the fallthrough path out of the AllocateArray may have become dead.
2597 // CatchNode::Value() is responsible for killing that path. The CatchNode has to be explicitly enqueued for igvn
2598 // to guarantee the change is not missed.
2599 if (use_op == Op_AllocateArray && n == use->in(AllocateNode::ValidLengthTest)) {
2600 Node* p = use->as_AllocateArray()->proj_out_or_null(TypeFunc::Control);
2601 if (p != nullptr) {
2602 add_users_to_worklist0(p, worklist);
2603 }
2604 }
2605
2606 if (use_op == Op_Initialize) {
2607 Node* imem = use->as_Initialize()->proj_out_or_null(TypeFunc::Memory);
2608 if (imem != nullptr) add_users_to_worklist0(imem, worklist);
2609 }
2610 // Loading the java mirror from a Klass requires two loads and the type
2611 // of the mirror load depends on the type of 'n'. See LoadNode::Value().
2612 // LoadBarrier?(LoadP(LoadP(AddP(foo:Klass, #java_mirror))))
2613 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2614 bool has_load_barrier_nodes = bs->has_load_barrier_nodes();
2615
2616 if (use_op == Op_CastP2X) {
2617 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2618 Node* u = use->fast_out(i2);
2619 if (u->Opcode() == Op_AndX) {
2620 worklist.push(u);
2621 }
2622 }
2623 }
2624 if (use_op == Op_LoadP && use->bottom_type()->isa_rawptr()) {
2625 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2626 Node* u = use->fast_out(i2);
2627 const Type* ut = u->bottom_type();
2628 if (u->Opcode() == Op_LoadP && ut->isa_instptr()) {
2629 if (has_load_barrier_nodes) {
2630 // Search for load barriers behind the load
2631 for (DUIterator_Fast i3max, i3 = u->fast_outs(i3max); i3 < i3max; i3++) {
2632 Node* b = u->fast_out(i3);
2633 if (bs->is_gc_barrier_node(b)) {
2634 worklist.push(b);
2635 }
2636 }
2637 }
2638 worklist.push(u);
2639 }
2640 }
2641 }
2642 // Give CallStaticJavaNode::remove_useless_allocation a chance to run
2643 if (use->is_Region()) {
2644 Node* c = use;
2645 do {
2646 c = c->unique_ctrl_out_or_null();
2647 } while (c != nullptr && c->is_Region());
2648 if (c != nullptr && c->is_CallStaticJava() && c->as_CallStaticJava()->uncommon_trap_request() != 0) {
2649 worklist.push(c);
2650 }
2651 }
2652 if (use->Opcode() == Op_OpaqueZeroTripGuard) {
2653 assert(use->outcnt() <= 1, "OpaqueZeroTripGuard can't be shared");
2654 if (use->outcnt() == 1) {
2655 Node* cmp = use->unique_out();
2656 worklist.push(cmp);
2657 }
2658 }
2659 if (use->Opcode() == Op_AddX) {
2660 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2661 Node* u = use->fast_out(i2);
2662 if (u->Opcode() == Op_CastX2P) {
2663 worklist.push(u);
2664 }
2665 }
2666 }
2667
2668 /* AndNode has a special handling when one of the operands is a LShiftNode:
2669 * (LHS << s) & RHS
2670 * if RHS fits in less than s bits, the value of this expression is 0.
2671 * The difficulty is that there might be a conversion node (ConvI2L) between
2672 * the LShiftINode and the AndLNode, like so:
2673 * AndLNode(ConvI2L(LShiftI(LHS, s)), RHS)
2674 * This case is handled by And[IL]Node::Value(PhaseGVN*)
2675 * (see `AndIL_min_trailing_zeros`).
2676 *
2677 * But, when the shift is updated during IGVN, pushing the user (ConvI2L)
2678 * is not enough: there might be no update happening there. We need to
2679 * directly push the And[IL]Node on the worklist, jumping over ConvI2L.
2680 *
2681 * Moreover we can have ConstraintCasts in between. It may look like
2682 * ConstraintCast+ -> ConvI2L -> ConstraintCast+ -> And
2683 * and And[IL]Node::Value(PhaseGVN*) still handles that by looking through casts.
2684 * So we must deal with that as well.
2685 */
2686 if (use->is_ConstraintCast() || use_op == Op_ConvI2L) {
2687 auto is_boundary = [](Node* n){ return !n->is_ConstraintCast() && n->Opcode() != Op_ConvI2L; };
2688 auto push_and_to_worklist = [&worklist](Node* n){
2689 if (n->Opcode() == Op_AndL || n->Opcode() == Op_AndI) {
2690 worklist.push(n);
2691 }
2692 };
2693 use->visit_uses(push_and_to_worklist, is_boundary);
2694 }
2695 }
2696
2697 /**
2698 * Remove the speculative part of all types that we know of
2699 */
2700 void PhaseIterGVN::remove_speculative_types() {
2701 assert(UseTypeSpeculation, "speculation is off");
2702 for (uint i = 0; i < _types.Size(); i++) {
2703 const Type* t = _types.fast_lookup(i);
2704 if (t != nullptr) {
2705 _types.map(i, t->remove_speculative());
2706 }
2707 }
2708 _table.check_no_speculative_types();
2709 }
2710
2711 // Check if the type of a divisor of a Div or Mod node includes zero.
2712 bool PhaseIterGVN::no_dependent_zero_check(Node* n) const {
2713 switch (n->Opcode()) {
2714 case Op_DivI:
2715 case Op_ModI:
2716 case Op_UDivI:
2717 case Op_UModI: {
2718 // Type of divisor includes 0?
2719 if (type(n->in(2)) == Type::TOP) {
2720 // 'n' is dead. Treat as if zero check is still there to avoid any further optimizations.
2721 return false;
2722 }
2723 const TypeInt* type_divisor = type(n->in(2))->is_int();
2724 return (type_divisor->_hi < 0 || type_divisor->_lo > 0);
2725 }
2726 case Op_DivL:
2727 case Op_ModL:
2728 case Op_UDivL:
2729 case Op_UModL: {
2730 // Type of divisor includes 0?
2731 if (type(n->in(2)) == Type::TOP) {
2732 // 'n' is dead. Treat as if zero check is still there to avoid any further optimizations.
2733 return false;
2734 }
2735 const TypeLong* type_divisor = type(n->in(2))->is_long();
2736 return (type_divisor->_hi < 0 || type_divisor->_lo > 0);
2737 }
2738 }
2739 return true;
2740 }
2741
2742 //=============================================================================
2743 #ifndef PRODUCT
2744 uint PhaseCCP::_total_invokes = 0;
2745 uint PhaseCCP::_total_constants = 0;
2746 #endif
2747 //------------------------------PhaseCCP---------------------------------------
2748 // Conditional Constant Propagation, ala Wegman & Zadeck
2749 PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) {
2750 NOT_PRODUCT( clear_constants(); )
2751 assert( _worklist.size() == 0, "" );
2752 analyze();
2753 }
2754
2755 #ifndef PRODUCT
2756 //------------------------------~PhaseCCP--------------------------------------
2757 PhaseCCP::~PhaseCCP() {
2758 inc_invokes();
2759 _total_constants += count_constants();
2760 }
2761 #endif
2762
2763
2764 #ifdef ASSERT
2765 void PhaseCCP::verify_type(Node* n, const Type* tnew, const Type* told) {
2766 if (tnew->meet(told) != tnew->remove_speculative()) {
2767 n->dump(3);
2768 tty->print("told = "); told->dump(); tty->cr();
2769 tty->print("tnew = "); tnew->dump(); tty->cr();
2770 fatal("Not monotonic");
2771 }
2772 assert(!told->isa_int() || !tnew->isa_int() || told->is_int()->_widen <= tnew->is_int()->_widen, "widen increases");
2773 assert(!told->isa_long() || !tnew->isa_long() || told->is_long()->_widen <= tnew->is_long()->_widen, "widen increases");
2774 }
2775 #endif //ASSERT
2776
2777 // In this analysis, all types are initially set to TOP. We iteratively call Value() on all nodes of the graph until
2778 // we reach a fixed-point (i.e. no types change anymore). We start with a list that only contains the root node. Each time
2779 // a new type is set, we push all uses of that node back to the worklist (in some cases, we also push grandchildren
2780 // or nodes even further down back to the worklist because their type could change as a result of the current type
2781 // change).
2782 void PhaseCCP::analyze() {
2783 // Initialize all types to TOP, optimistic analysis
2784 for (uint i = 0; i < C->unique(); i++) {
2785 _types.map(i, Type::TOP);
2786 }
2787
2788 // CCP worklist is placed on a local arena, so that we can allow ResourceMarks on "Compile::current()->resource_arena()".
2789 // We also do not want to put the worklist on "Compile::current()->comp_arena()", as that one only gets de-allocated after
2790 // Compile is over. The local arena gets de-allocated at the end of its scope.
2791 ResourceArea local_arena(mtCompiler);
2792 Unique_Node_List worklist(&local_arena);
2793 DEBUG_ONLY(Unique_Node_List worklist_verify(&local_arena);)
2794
2795 // Push root onto worklist
2796 worklist.push(C->root());
2797
2798 assert(_root_and_safepoints.size() == 0, "must be empty (unused)");
2799 _root_and_safepoints.push(C->root());
2800
2801 // Pull from worklist; compute new value; push changes out.
2802 // This loop is the meat of CCP.
2803 while (worklist.size() != 0) {
2804 Node* n = fetch_next_node(worklist);
2805 DEBUG_ONLY(worklist_verify.push(n);)
2806 if (n->is_SafePoint()) {
2807 // Make sure safepoints are processed by PhaseCCP::transform even if they are
2808 // not reachable from the bottom. Otherwise, infinite loops would be removed.
2809 _root_and_safepoints.push(n);
2810 }
2811 const Type* new_type = n->Value(this);
2812 if (new_type != type(n)) {
2813 DEBUG_ONLY(verify_type(n, new_type, type(n));)
2814 dump_type_and_node(n, new_type);
2815 set_type(n, new_type);
2816 push_child_nodes_to_worklist(worklist, n);
2817 }
2818 if (KillPathsReachableByDeadTypeNode && n->is_Type() && new_type == Type::TOP) {
2819 // Keep track of Type nodes to kill CFG paths that use Type
2820 // nodes that become dead.
2821 _maybe_top_type_nodes.push(n);
2822 }
2823 }
2824 DEBUG_ONLY(verify_analyze(worklist_verify);)
2825 }
2826
2827 #ifdef ASSERT
2828 // For every node n on verify list, check if type(n) == n->Value()
2829 // We have a list of exceptions, see comments in verify_Value_for.
2830 void PhaseCCP::verify_analyze(Unique_Node_List& worklist_verify) {
2831 bool failure = false;
2832 while (worklist_verify.size()) {
2833 Node* n = worklist_verify.pop();
2834 failure |= verify_Value_for(n);
2835 }
2836 // If we get this assert, check why the reported nodes were not processed again in CCP.
2837 // We should either make sure that these nodes are properly added back to the CCP worklist
2838 // in PhaseCCP::push_child_nodes_to_worklist() to update their type or add an exception
2839 // in the verification code above if that is not possible for some reason (like Load nodes).
2840 assert(!failure, "PhaseCCP not at fixpoint: analysis result may be unsound.");
2841 }
2842 #endif
2843
2844 // Fetch next node from worklist to be examined in this iteration.
2845 Node* PhaseCCP::fetch_next_node(Unique_Node_List& worklist) {
2846 if (StressCCP) {
2847 return worklist.remove(C->random() % worklist.size());
2848 } else {
2849 return worklist.pop();
2850 }
2851 }
2852
2853 #ifndef PRODUCT
2854 void PhaseCCP::dump_type_and_node(const Node* n, const Type* t) {
2855 if (TracePhaseCCP) {
2856 t->dump();
2857 do {
2858 tty->print("\t");
2859 } while (tty->position() < 16);
2860 n->dump();
2861 }
2862 }
2863 #endif
2864
2865 // We need to propagate the type change of 'n' to all its uses. Depending on the kind of node, additional nodes
2866 // (grandchildren or even further down) need to be revisited as their types could also be improved as a result
2867 // of the new type of 'n'. Push these nodes to the worklist.
2868 void PhaseCCP::push_child_nodes_to_worklist(Unique_Node_List& worklist, Node* n) const {
2869 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2870 Node* use = n->fast_out(i);
2871 push_if_not_bottom_type(worklist, use);
2872 push_more_uses(worklist, n, use);
2873 }
2874 }
2875
2876 void PhaseCCP::push_if_not_bottom_type(Unique_Node_List& worklist, Node* n) const {
2877 if (n->bottom_type() != type(n)) {
2878 worklist.push(n);
2879 }
2880 }
2881
2882 // For some nodes, we need to propagate the type change to grandchildren or even further down.
2883 // Add them back to the worklist.
2884 void PhaseCCP::push_more_uses(Unique_Node_List& worklist, Node* parent, const Node* use) const {
2885 push_phis(worklist, use);
2886 push_catch(worklist, use);
2887 push_cmpu(worklist, use);
2888 push_counted_loop_phi(worklist, parent, use);
2889 push_cast(worklist, use);
2890 push_loadp(worklist, use);
2891 push_and(worklist, parent, use);
2892 push_cast_ii(worklist, parent, use);
2893 push_opaque_zero_trip_guard(worklist, use);
2894 }
2895
2896
2897 // We must recheck Phis too if use is a Region.
2898 void PhaseCCP::push_phis(Unique_Node_List& worklist, const Node* use) const {
2899 if (use->is_Region()) {
2900 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
2901 push_if_not_bottom_type(worklist, use->fast_out(i));
2902 }
2903 }
2904 }
2905
2906 // If we changed the receiver type to a call, we need to revisit the Catch node following the call. It's looking for a
2907 // non-null receiver to know when to enable the regular fall-through path in addition to the NullPtrException path.
2908 // Same is true if the type of a ValidLengthTest input to an AllocateArrayNode changes.
2909 void PhaseCCP::push_catch(Unique_Node_List& worklist, const Node* use) {
2910 if (use->is_Call()) {
2911 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
2912 Node* proj = use->fast_out(i);
2913 if (proj->is_Proj() && proj->as_Proj()->_con == TypeFunc::Control) {
2914 Node* catch_node = proj->find_out_with(Op_Catch);
2915 if (catch_node != nullptr) {
2916 worklist.push(catch_node);
2917 }
2918 }
2919 }
2920 }
2921 }
2922
2923 // CmpU nodes can get their type information from two nodes up in the graph (instead of from the nodes immediately
2924 // above). Make sure they are added to the worklist if nodes they depend on are updated since they could be missed
2925 // and get wrong types otherwise.
2926 void PhaseCCP::push_cmpu(Unique_Node_List& worklist, const Node* use) const {
2927 uint use_op = use->Opcode();
2928 if (use_op == Op_AddI || use_op == Op_SubI) {
2929 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
2930 Node* cmpu = use->fast_out(i);
2931 const uint cmpu_opcode = cmpu->Opcode();
2932 if (cmpu_opcode == Op_CmpU || cmpu_opcode == Op_CmpU3) {
2933 // Got a CmpU or CmpU3 which might need the new type information from node n.
2934 push_if_not_bottom_type(worklist, cmpu);
2935 }
2936 }
2937 }
2938 }
2939
2940 // If n is used in a counted loop exit condition, then the type of the counted loop's Phi depends on the type of 'n'.
2941 // Seem PhiNode::Value().
2942 void PhaseCCP::push_counted_loop_phi(Unique_Node_List& worklist, Node* parent, const Node* use) {
2943 uint use_op = use->Opcode();
2944 if (use_op == Op_CmpI || use_op == Op_CmpL) {
2945 PhiNode* phi = countedloop_phi_from_cmp(use->as_Cmp(), parent);
2946 if (phi != nullptr) {
2947 worklist.push(phi);
2948 }
2949 }
2950 }
2951
2952 void PhaseCCP::push_cast(Unique_Node_List& worklist, const Node* use) {
2953 uint use_op = use->Opcode();
2954 if (use_op == Op_CastP2X) {
2955 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
2956 Node* u = use->fast_out(i2);
2957 if (u->Opcode() == Op_AndX) {
2958 worklist.push(u);
2959 }
2960 }
2961 }
2962 }
2963
2964 // Loading the java mirror from a Klass requires two loads and the type of the mirror load depends on the type of 'n'.
2965 // See LoadNode::Value().
2966 void PhaseCCP::push_loadp(Unique_Node_List& worklist, const Node* use) const {
2967 BarrierSetC2* barrier_set = BarrierSet::barrier_set()->barrier_set_c2();
2968 bool has_load_barrier_nodes = barrier_set->has_load_barrier_nodes();
2969
2970 if (use->Opcode() == Op_LoadP && use->bottom_type()->isa_rawptr()) {
2971 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
2972 Node* loadp = use->fast_out(i);
2973 const Type* ut = loadp->bottom_type();
2974 if (loadp->Opcode() == Op_LoadP && ut->isa_instptr() && ut != type(loadp)) {
2975 if (has_load_barrier_nodes) {
2976 // Search for load barriers behind the load
2977 push_load_barrier(worklist, barrier_set, loadp);
2978 }
2979 worklist.push(loadp);
2980 }
2981 }
2982 }
2983 }
2984
2985 void PhaseCCP::push_load_barrier(Unique_Node_List& worklist, const BarrierSetC2* barrier_set, const Node* use) {
2986 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
2987 Node* barrier_node = use->fast_out(i);
2988 if (barrier_set->is_gc_barrier_node(barrier_node)) {
2989 worklist.push(barrier_node);
2990 }
2991 }
2992 }
2993
2994 // AndI/L::Value() optimizes patterns similar to (v << 2) & 3, or CON & 3 to zero if they are bitwise disjoint.
2995 // Add the AndI/L nodes back to the worklist to re-apply Value() in case the value is now a constant or shift
2996 // value changed.
2997 void PhaseCCP::push_and(Unique_Node_List& worklist, const Node* parent, const Node* use) const {
2998 const TypeInteger* parent_type = type(parent)->isa_integer(type(parent)->basic_type());
2999 uint use_op = use->Opcode();
3000 if (
3001 // Pattern: parent (now constant) -> (ConstraintCast | ConvI2L)* -> And
3002 (parent_type != nullptr && parent_type->is_con()) ||
3003 // Pattern: parent -> LShift (use) -> (ConstraintCast | ConvI2L)* -> And
3004 ((use_op == Op_LShiftI || use_op == Op_LShiftL) && use->in(2) == parent)) {
3005
3006 auto push_and_uses_to_worklist = [&](Node* n) {
3007 uint opc = n->Opcode();
3008 if (opc == Op_AndI || opc == Op_AndL) {
3009 push_if_not_bottom_type(worklist, n);
3010 }
3011 };
3012 auto is_boundary = [](Node* n) {
3013 return !(n->is_ConstraintCast() || n->Opcode() == Op_ConvI2L);
3014 };
3015 use->visit_uses(push_and_uses_to_worklist, is_boundary);
3016 }
3017 }
3018
3019 // CastII::Value() optimizes CmpI/If patterns if the right input of the CmpI has a constant type. If the CastII input is
3020 // the same node as the left input into the CmpI node, the type of the CastII node can be improved accordingly. Add the
3021 // CastII node back to the worklist to re-apply Value() to either not miss this optimization or to undo it because it
3022 // cannot be applied anymore. We could have optimized the type of the CastII before but now the type of the right input
3023 // of the CmpI (i.e. 'parent') is no longer constant. The type of the CastII must be widened in this case.
3024 void PhaseCCP::push_cast_ii(Unique_Node_List& worklist, const Node* parent, const Node* use) const {
3025 if (use->Opcode() == Op_CmpI && use->in(2) == parent) {
3026 Node* other_cmp_input = use->in(1);
3027 for (DUIterator_Fast imax, i = other_cmp_input->fast_outs(imax); i < imax; i++) {
3028 Node* cast_ii = other_cmp_input->fast_out(i);
3029 if (cast_ii->is_CastII()) {
3030 push_if_not_bottom_type(worklist, cast_ii);
3031 }
3032 }
3033 }
3034 }
3035
3036 void PhaseCCP::push_opaque_zero_trip_guard(Unique_Node_List& worklist, const Node* use) const {
3037 if (use->Opcode() == Op_OpaqueZeroTripGuard) {
3038 push_if_not_bottom_type(worklist, use->unique_out());
3039 }
3040 }
3041
3042 //------------------------------do_transform-----------------------------------
3043 // Top level driver for the recursive transformer
3044 void PhaseCCP::do_transform() {
3045 // Correct leaves of new-space Nodes; they point to old-space.
3046 C->set_root( transform(C->root())->as_Root() );
3047 assert( C->top(), "missing TOP node" );
3048 assert( C->root(), "missing root" );
3049 }
3050
3051 //------------------------------transform--------------------------------------
3052 // Given a Node in old-space, clone him into new-space.
3053 // Convert any of his old-space children into new-space children.
3054 Node *PhaseCCP::transform( Node *n ) {
3055 assert(n->is_Root(), "traversal must start at root");
3056 assert(_root_and_safepoints.member(n), "root (n) must be in list");
3057
3058 ResourceMark rm;
3059 // Map: old node idx -> node after CCP (or nullptr if not yet transformed or useless).
3060 Node_List node_map;
3061 // Pre-allocate to avoid frequent realloc
3062 GrowableArray <Node *> transform_stack(C->live_nodes() >> 1);
3063 // track all visited nodes, so that we can remove the complement
3064 Unique_Node_List useful;
3065
3066 if (KillPathsReachableByDeadTypeNode) {
3067 for (uint i = 0; i < _maybe_top_type_nodes.size(); ++i) {
3068 Node* type_node = _maybe_top_type_nodes.at(i);
3069 if (type(type_node) == Type::TOP) {
3070 ResourceMark rm;
3071 type_node->as_Type()->make_paths_from_here_dead(this, nullptr, "ccp");
3072 }
3073 }
3074 } else {
3075 assert(_maybe_top_type_nodes.size() == 0, "we don't need type nodes");
3076 }
3077
3078 // Initialize the traversal.
3079 // This CCP pass may prove that no exit test for a loop ever succeeds (i.e. the loop is infinite). In that case,
3080 // the logic below doesn't follow any path from Root to the loop body: there's at least one such path but it's proven
3081 // never taken (its type is TOP). As a consequence the node on the exit path that's input to Root (let's call it n) is
3082 // replaced by the top node and the inputs of that node n are not enqueued for further processing. If CCP only works
3083 // through the graph from Root, this causes the loop body to never be processed here even when it's not dead (that
3084 // is reachable from Root following its uses). To prevent that issue, transform() starts walking the graph from Root
3085 // and all safepoints.
3086 for (uint i = 0; i < _root_and_safepoints.size(); ++i) {
3087 Node* nn = _root_and_safepoints.at(i);
3088 Node* new_node = node_map[nn->_idx];
3089 assert(new_node == nullptr, "");
3090 new_node = transform_once(nn); // Check for constant
3091 node_map.map(nn->_idx, new_node); // Flag as having been cloned
3092 transform_stack.push(new_node); // Process children of cloned node
3093 useful.push(new_node);
3094 }
3095
3096 while (transform_stack.is_nonempty()) {
3097 Node* clone = transform_stack.pop();
3098 uint cnt = clone->req();
3099 for( uint i = 0; i < cnt; i++ ) { // For all inputs do
3100 Node *input = clone->in(i);
3101 if( input != nullptr ) { // Ignore nulls
3102 Node *new_input = node_map[input->_idx]; // Check for cloned input node
3103 if( new_input == nullptr ) {
3104 new_input = transform_once(input); // Check for constant
3105 node_map.map( input->_idx, new_input );// Flag as having been cloned
3106 transform_stack.push(new_input); // Process children of cloned node
3107 useful.push(new_input);
3108 }
3109 assert( new_input == clone->in(i), "insanity check");
3110 }
3111 }
3112 }
3113
3114 // The above transformation might lead to subgraphs becoming unreachable from the
3115 // bottom while still being reachable from the top. As a result, nodes in that
3116 // subgraph are not transformed and their bottom types are not updated, leading to
3117 // an inconsistency between bottom_type() and type(). In rare cases, LoadNodes in
3118 // such a subgraph, might be re-enqueued for IGVN indefinitely by MemNode::Ideal_common
3119 // because their address type is inconsistent. Therefore, we aggressively remove
3120 // all useless nodes here even before PhaseIdealLoop::build_loop_late gets a chance
3121 // to remove them anyway.
3122 if (C->cached_top_node()) {
3123 useful.push(C->cached_top_node());
3124 }
3125 C->update_dead_node_list(useful);
3126 remove_useless_nodes(useful.member_set());
3127 _worklist.remove_useless_nodes(useful.member_set());
3128 C->disconnect_useless_nodes(useful, _worklist, &_root_and_safepoints);
3129
3130 Node* new_root = node_map[n->_idx];
3131 assert(new_root->is_Root(), "transformed root node must be a root node");
3132 return new_root;
3133 }
3134
3135 //------------------------------transform_once---------------------------------
3136 // For PhaseCCP, transformation is IDENTITY unless Node computed a constant.
3137 Node *PhaseCCP::transform_once( Node *n ) {
3138 const Type *t = type(n);
3139 // Constant? Use constant Node instead
3140 if( t->singleton() ) {
3141 Node *nn = n; // Default is to return the original constant
3142 if( t == Type::TOP ) {
3143 // cache my top node on the Compile instance
3144 if( C->cached_top_node() == nullptr || C->cached_top_node()->in(0) == nullptr ) {
3145 C->set_cached_top_node(ConNode::make(Type::TOP));
3146 set_type(C->top(), Type::TOP);
3147 }
3148 nn = C->top();
3149 }
3150 if( !n->is_Con() ) {
3151 if( t != Type::TOP ) {
3152 nn = makecon(t); // ConNode::make(t);
3153 NOT_PRODUCT( inc_constants(); )
3154 } else if( n->is_Region() ) { // Unreachable region
3155 // Note: nn == C->top()
3156 n->set_req(0, nullptr); // Cut selfreference
3157 bool progress = true;
3158 uint max = n->outcnt();
3159 DUIterator i;
3160 while (progress) {
3161 progress = false;
3162 // Eagerly remove dead phis to avoid phis copies creation.
3163 for (i = n->outs(); n->has_out(i); i++) {
3164 Node* m = n->out(i);
3165 if (m->is_Phi()) {
3166 assert(type(m) == Type::TOP, "Unreachable region should not have live phis.");
3167 replace_node(m, nn);
3168 if (max != n->outcnt()) {
3169 progress = true;
3170 i = n->refresh_out_pos(i);
3171 max = n->outcnt();
3172 }
3173 }
3174 }
3175 }
3176 }
3177 replace_node(n,nn); // Update DefUse edges for new constant
3178 }
3179 return nn;
3180 }
3181
3182 // If x is a TypeNode, capture any more-precise type permanently into Node
3183 if (t != n->bottom_type()) {
3184 hash_delete(n); // changing bottom type may force a rehash
3185 n->raise_bottom_type(t);
3186 _worklist.push(n); // n re-enters the hash table via the worklist
3187 add_users_to_worklist(n); // if ideal or identity optimizations depend on the input type, users need to be notified
3188 }
3189
3190 // TEMPORARY fix to ensure that 2nd GVN pass eliminates null checks
3191 switch( n->Opcode() ) {
3192 case Op_CallStaticJava: // Give post-parse call devirtualization a chance
3193 case Op_CallDynamicJava:
3194 case Op_FastLock: // Revisit FastLocks for lock coarsening
3195 case Op_If:
3196 case Op_CountedLoopEnd:
3197 case Op_Region:
3198 case Op_Loop:
3199 case Op_CountedLoop:
3200 case Op_Conv2B:
3201 case Op_Opaque1:
3202 _worklist.push(n);
3203 break;
3204 default:
3205 break;
3206 }
3207
3208 return n;
3209 }
3210
3211 //---------------------------------saturate------------------------------------
3212 const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type,
3213 const Type* limit_type) const {
3214 const Type* wide_type = new_type->widen(old_type, limit_type);
3215 if (wide_type != new_type) { // did we widen?
3216 // If so, we may have widened beyond the limit type. Clip it back down.
3217 new_type = wide_type->filter(limit_type);
3218 }
3219 return new_type;
3220 }
3221
3222 //------------------------------print_statistics-------------------------------
3223 #ifndef PRODUCT
3224 void PhaseCCP::print_statistics() {
3225 tty->print_cr("CCP: %d constants found: %d", _total_invokes, _total_constants);
3226 }
3227 #endif
3228
3229
3230 //=============================================================================
3231 #ifndef PRODUCT
3232 uint PhasePeephole::_total_peepholes = 0;
3233 #endif
3234 //------------------------------PhasePeephole----------------------------------
3235 // Conditional Constant Propagation, ala Wegman & Zadeck
3236 PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg )
3237 : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) {
3238 NOT_PRODUCT( clear_peepholes(); )
3239 }
3240
3241 #ifndef PRODUCT
3242 //------------------------------~PhasePeephole---------------------------------
3243 PhasePeephole::~PhasePeephole() {
3244 _total_peepholes += count_peepholes();
3245 }
3246 #endif
3247
3248 //------------------------------transform--------------------------------------
3249 Node *PhasePeephole::transform( Node *n ) {
3250 ShouldNotCallThis();
3251 return nullptr;
3252 }
3253
3254 //------------------------------do_transform-----------------------------------
3255 void PhasePeephole::do_transform() {
3256 bool method_name_not_printed = true;
3257
3258 // Examine each basic block
3259 for (uint block_number = 1; block_number < _cfg.number_of_blocks(); ++block_number) {
3260 Block* block = _cfg.get_block(block_number);
3261 bool block_not_printed = true;
3262
3263 for (bool progress = true; progress;) {
3264 progress = false;
3265 // block->end_idx() not valid after PhaseRegAlloc
3266 uint end_index = block->number_of_nodes();
3267 for( uint instruction_index = end_index - 1; instruction_index > 0; --instruction_index ) {
3268 Node *n = block->get_node(instruction_index);
3269 if( n->is_Mach() ) {
3270 MachNode *m = n->as_Mach();
3271 // check for peephole opportunities
3272 int result = m->peephole(block, instruction_index, &_cfg, _regalloc);
3273 if( result != -1 ) {
3274 #ifndef PRODUCT
3275 if( PrintOptoPeephole ) {
3276 // Print method, first time only
3277 if( C->method() && method_name_not_printed ) {
3278 C->method()->print_short_name(); tty->cr();
3279 method_name_not_printed = false;
3280 }
3281 // Print this block
3282 if( Verbose && block_not_printed) {
3283 tty->print_cr("in block");
3284 block->dump();
3285 block_not_printed = false;
3286 }
3287 // Print the peephole number
3288 tty->print_cr("peephole number: %d", result);
3289 }
3290 inc_peepholes();
3291 #endif
3292 // Set progress, start again
3293 progress = true;
3294 break;
3295 }
3296 }
3297 }
3298 }
3299 }
3300 }
3301
3302 //------------------------------print_statistics-------------------------------
3303 #ifndef PRODUCT
3304 void PhasePeephole::print_statistics() {
3305 tty->print_cr("Peephole: peephole rules applied: %d", _total_peepholes);
3306 }
3307 #endif
3308
3309
3310 //=============================================================================
3311 //------------------------------set_req_X--------------------------------------
3312 void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) {
3313 assert( is_not_dead(n), "can not use dead node");
3314 #ifdef ASSERT
3315 if (igvn->hash_find(this) == this) {
3316 tty->print_cr("Need to remove from hash before changing edges");
3317 this->dump(1);
3318 tty->print_cr("Set at i = %d", i);
3319 n->dump();
3320 assert(false, "Need to remove from hash before changing edges");
3321 }
3322 #endif
3323 Node *old = in(i);
3324 set_req(i, n);
3325
3326 // old goes dead?
3327 if( old ) {
3328 switch (old->outcnt()) {
3329 case 0:
3330 // Put into the worklist to kill later. We do not kill it now because the
3331 // recursive kill will delete the current node (this) if dead-loop exists
3332 if (!old->is_top())
3333 igvn->_worklist.push( old );
3334 break;
3335 case 1:
3336 if( old->is_Store() || old->has_special_unique_user() )
3337 igvn->add_users_to_worklist( old );
3338 break;
3339 case 2:
3340 if( old->is_Store() )
3341 igvn->add_users_to_worklist( old );
3342 if( old->Opcode() == Op_Region )
3343 igvn->_worklist.push(old);
3344 break;
3345 case 3:
3346 if( old->Opcode() == Op_Region ) {
3347 igvn->_worklist.push(old);
3348 igvn->add_users_to_worklist( old );
3349 }
3350 break;
3351 default:
3352 break;
3353 }
3354
3355 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(igvn, old);
3356 }
3357 }
3358
3359 void Node::set_req_X(uint i, Node *n, PhaseGVN *gvn) {
3360 PhaseIterGVN* igvn = gvn->is_IterGVN();
3361 if (igvn == nullptr) {
3362 set_req(i, n);
3363 return;
3364 }
3365 set_req_X(i, n, igvn);
3366 }
3367
3368 //-------------------------------replace_by-----------------------------------
3369 // Using def-use info, replace one node for another. Follow the def-use info
3370 // to all users of the OLD node. Then make all uses point to the NEW node.
3371 void Node::replace_by(Node *new_node) {
3372 assert(!is_top(), "top node has no DU info");
3373 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) {
3374 Node* use = last_out(i);
3375 uint uses_found = 0;
3376 for (uint j = 0; j < use->len(); j++) {
3377 if (use->in(j) == this) {
3378 if (j < use->req())
3379 use->set_req(j, new_node);
3380 else use->set_prec(j, new_node);
3381 uses_found++;
3382 }
3383 }
3384 i -= uses_found; // we deleted 1 or more copies of this edge
3385 }
3386 }
3387
3388 //=============================================================================
3389 //-----------------------------------------------------------------------------
3390 void Type_Array::grow( uint i ) {
3391 assert(_a == Compile::current()->comp_arena(), "Should be allocated in comp_arena");
3392 if( !_max ) {
3393 _max = 1;
3394 _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) );
3395 _types[0] = nullptr;
3396 }
3397 uint old = _max;
3398 _max = next_power_of_2(i);
3399 _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*));
3400 memset( &_types[old], 0, (_max-old)*sizeof(Type*) );
3401 }
3402
3403 //------------------------------dump-------------------------------------------
3404 #ifndef PRODUCT
3405 void Type_Array::dump() const {
3406 uint max = Size();
3407 for( uint i = 0; i < max; i++ ) {
3408 if( _types[i] != nullptr ) {
3409 tty->print(" %d\t== ", i); _types[i]->dump(); tty->cr();
3410 }
3411 }
3412 }
3413 #endif